Merge pull request 'feat(social): speech + LLM + TTS + orchestrator (#81 #83 #85 #89)' (#102) from sl-jetson/social-speech-llm-tts into main

jetson/ros2_ws/src/saltybot_social/config/conversation_params.yaml

@@ -0,0 +1,12 @@
+conversation_node:
+  ros__parameters:
+    model_path: "/models/gguf/phi-3-mini-4k-instruct-q4_k_m.gguf"
+    n_ctx: 4096
+    n_gpu_layers: 20              # Increase for more VRAM (Orin has 8GB shared)
+    max_tokens: 200
+    temperature: 0.7
+    top_p: 0.9
+    soul_path: "/soul/SOUL.md"
+    context_db_path: "/social_db/conversation_context.json"
+    save_interval_s: 30.0
+    stream: true

jetson/ros2_ws/src/saltybot_social/config/orchestrator_params.yaml

@@ -0,0 +1,8 @@
+orchestrator_node:
+  ros__parameters:
+    gpu_mem_warn_mb: 4000.0       # Warn when GPU free < 4GB
+    gpu_mem_throttle_mb: 2000.0   # Throttle new inferences at < 2GB
+    watchdog_timeout_s: 30.0
+    latency_window: 20
+    profile_enabled: true
+    state_publish_rate: 2.0       # Hz

jetson/ros2_ws/src/saltybot_social/config/speech_params.yaml

@@ -0,0 +1,13 @@
+speech_pipeline_node:
+  ros__parameters:
+    mic_device_index: -1          # -1 = system default; use `arecord -l` to list
+    sample_rate: 16000
+    wake_word_model: "hey_salty"
+    wake_word_threshold: 0.5
+    vad_threshold_db: -35.0
+    use_silero_vad: true
+    whisper_model: "small"        # small (~500ms), medium (better quality, ~900ms)
+    whisper_compute_type: "float16"
+    speaker_threshold: 0.65
+    speaker_db_path: "/social_db/speaker_embeddings.json"
+    publish_partial: true

jetson/ros2_ws/src/saltybot_social/config/tts_params.yaml

@@ -0,0 +1,9 @@
+tts_node:
+  ros__parameters:
+    voice_path: "/models/piper/en_US-lessac-medium.onnx"
+    sample_rate: 22050
+    volume: 1.0
+    audio_device: ""              # "" = system default; set to device name if needed
+    playback_enabled: true
+    publish_audio: false
+    sentence_streaming: true

jetson/ros2_ws/src/saltybot_social/launch/social_bot.launch.py

@@ -0,0 +1,129 @@
+"""social_bot.launch.py — Launch the full social-bot pipeline.
+
+Issue #89: End-to-end pipeline orchestrator launch file.
+
+Brings up all social-bot nodes in dependency order:
+  1. Orchestrator (state machine + watchdog)  — t=0s
+  2. Speech pipeline (wake word + VAD + STT)  — t=2s
+  3. Conversation engine (LLM)                — t=4s
+  4. TTS node (Piper streaming)               — t=4s
+  5. Person state tracker (already in social.launch.py — optional)
+
+Launch args:
+  enable_speech       (bool, true)
+  enable_llm          (bool, true)
+  enable_tts          (bool, true)
+  enable_orchestrator (bool, true)
+  voice_path          (str, /models/piper/en_US-lessac-medium.onnx)
+  whisper_model       (str, small)
+  llm_model_path      (str, /models/gguf/phi-3-mini-4k-instruct-q4_k_m.gguf)
+  n_gpu_layers        (int, 20)
+"""
+
+import os
+from ament_index_python.packages import get_package_share_directory
+from launch import LaunchDescription
+from launch.actions import (
+    DeclareLaunchArgument, GroupAction, LogInfo, TimerAction
+)
+from launch.conditions import IfCondition
+from launch.substitutions import LaunchConfiguration, PythonExpression
+from launch_ros.actions import Node
+
+
+def generate_launch_description() -> LaunchDescription:
+    pkg = get_package_share_directory("saltybot_social")
+    cfg = os.path.join(pkg, "config")
+
+    # ── Launch arguments ─────────────────────────────────────────────────────
+    args = [
+        DeclareLaunchArgument("enable_speech",       default_value="true"),
+        DeclareLaunchArgument("enable_llm",          default_value="true"),
+        DeclareLaunchArgument("enable_tts",          default_value="true"),
+        DeclareLaunchArgument("enable_orchestrator", default_value="true"),
+        DeclareLaunchArgument("voice_path",
+                              default_value="/models/piper/en_US-lessac-medium.onnx"),
+        DeclareLaunchArgument("whisper_model",       default_value="small"),
+        DeclareLaunchArgument("llm_model_path",
+                              default_value="/models/gguf/phi-3-mini-4k-instruct-q4_k_m.gguf"),
+        DeclareLaunchArgument("n_gpu_layers",        default_value="20"),
+    ]
+
+    # ── Orchestrator (t=0s) ──────────────────────────────────────────────────
+    orchestrator = Node(
+        package="saltybot_social",
+        executable="orchestrator_node",
+        name="orchestrator_node",
+        parameters=[os.path.join(cfg, "orchestrator_params.yaml")],
+        condition=IfCondition(LaunchConfiguration("enable_orchestrator")),
+        output="screen",
+        emulate_tty=True,
+    )
+
+    # ── Speech pipeline (t=2s) ───────────────────────────────────────────────
+    speech = TimerAction(period=2.0, actions=[
+        GroupAction(
+            condition=IfCondition(LaunchConfiguration("enable_speech")),
+            actions=[
+                LogInfo(msg="[social_bot] Starting speech pipeline..."),
+                Node(
+                    package="saltybot_social",
+                    executable="speech_pipeline_node",
+                    name="speech_pipeline_node",
+                    parameters=[
+                        os.path.join(cfg, "speech_params.yaml"),
+                        {"whisper_model": LaunchConfiguration("whisper_model")},
+                    ],
+                    output="screen",
+                    emulate_tty=True,
+                ),
+            ]
+        )
+    ])
+
+    # ── LLM conversation engine (t=4s) ───────────────────────────────────────
+    llm = TimerAction(period=4.0, actions=[
+        GroupAction(
+            condition=IfCondition(LaunchConfiguration("enable_llm")),
+            actions=[
+                LogInfo(msg="[social_bot] Starting LLM conversation engine..."),
+                Node(
+                    package="saltybot_social",
+                    executable="conversation_node",
+                    name="conversation_node",
+                    parameters=[
+                        os.path.join(cfg, "conversation_params.yaml"),
+                        {
+                            "model_path": LaunchConfiguration("llm_model_path"),
+                            "n_gpu_layers": LaunchConfiguration("n_gpu_layers"),
+                        },
+                    ],
+                    output="screen",
+                    emulate_tty=True,
+                ),
+            ]
+        )
+    ])
+
+    # ── TTS node (t=4s, parallel with LLM) ──────────────────────────────────
+    tts = TimerAction(period=4.0, actions=[
+        GroupAction(
+            condition=IfCondition(LaunchConfiguration("enable_tts")),
+            actions=[
+                LogInfo(msg="[social_bot] Starting TTS node..."),
+                Node(
+                    package="saltybot_social",
+                    executable="tts_node",
+                    name="tts_node",
+                    parameters=[
+                        os.path.join(cfg, "tts_params.yaml"),
+                        {"voice_path": LaunchConfiguration("voice_path")},
+                    ],
+                    output="screen",
+                    emulate_tty=True,
+                ),
+            ]
+        )
+    ])
+
+    return LaunchDescription(args + [orchestrator, speech, llm, tts])

jetson/ros2_ws/src/saltybot_social/package.xml

@@ -5,9 +5,12 @@
   <version>0.1.0</version>
   <description>
     Social interaction layer for saltybot.
+    speech_pipeline_node: wake word + VAD + Whisper STT + diarization (Issue #81).
+    conversation_node: local LLM with per-person context (Issue #83).
+    tts_node: streaming TTS with Piper first-chunk (Issue #85).
+    orchestrator_node: pipeline state machine + GPU watchdog + latency profiler (Issue #89).
     person_state_tracker: multi-modal person identity fusion (Issue #82).
-    expression_node: bridges /social/mood to ESP32-C3 NeoPixel ring over serial (Issue #86).
+    expression_node: LED expression + motor attention (Issue #86).
-    attention_node: rotates robot toward active speaker via /social/persons bearing (Issue #86).
   </description>
   <maintainer email="seb@vayrette.com">seb</maintainer>
   <license>MIT</license>

jetson/ros2_ws/src/saltybot_social/saltybot_social/conversation_node.py

@@ -0,0 +1,267 @@
+"""conversation_node.py — Local LLM conversation engine with per-person context.
+
+Issue #83: Conversation engine for social-bot.
+
+Stack: Phi-3-mini or Llama-3.2-3B GGUF Q4_K_M via llama-cpp-python (CUDA).
+Subscribes /social/speech/transcript → builds per-person prompt → streams
+token output → publishes /social/conversation/response.
+
+Streaming: publishes partial=true tokens as they arrive, then final=false
+at end of generation. TTS node can begin synthesis on first sentence boundary.
+
+ROS2 topics:
+  Subscribe:  /social/speech/transcript (saltybot_social_msgs/SpeechTranscript)
+  Publish:    /social/conversation/response (saltybot_social_msgs/ConversationResponse)
+
+Parameters:
+  model_path       (str, "/models/gguf/phi-3-mini-4k-instruct-q4_k_m.gguf")
+  n_ctx            (int, 4096)
+  n_gpu_layers     (int, 20)    — GPU offload layers (increase for more VRAM usage)
+  max_tokens       (int, 200)
+  temperature      (float, 0.7)
+  top_p            (float, 0.9)
+  soul_path        (str, "/soul/SOUL.md")
+  context_db_path  (str, "/social_db/conversation_context.json")
+  save_interval_s  (float, 30.0) — how often to persist context to disk
+  stream           (bool, true)
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+from typing import Optional
+
+import rclpy
+from rclpy.node import Node
+from rclpy.qos import QoSProfile
+
+from saltybot_social_msgs.msg import SpeechTranscript, ConversationResponse
+from .llm_context import ContextStore, build_llama_prompt, load_system_prompt, needs_summary_prompt
+
+
+class ConversationNode(Node):
+    """Local LLM inference node with per-person conversation memory."""
+
+    def __init__(self) -> None:
+        super().__init__("conversation_node")
+
+        # ── Parameters ──────────────────────────────────────────────────────
+        self.declare_parameter("model_path",
+                               "/models/gguf/phi-3-mini-4k-instruct-q4_k_m.gguf")
+        self.declare_parameter("n_ctx", 4096)
+        self.declare_parameter("n_gpu_layers", 20)
+        self.declare_parameter("max_tokens", 200)
+        self.declare_parameter("temperature", 0.7)
+        self.declare_parameter("top_p", 0.9)
+        self.declare_parameter("soul_path", "/soul/SOUL.md")
+        self.declare_parameter("context_db_path", "/social_db/conversation_context.json")
+        self.declare_parameter("save_interval_s", 30.0)
+        self.declare_parameter("stream", True)
+
+        self._model_path = self.get_parameter("model_path").value
+        self._n_ctx = self.get_parameter("n_ctx").value
+        self._n_gpu = self.get_parameter("n_gpu_layers").value
+        self._max_tokens = self.get_parameter("max_tokens").value
+        self._temperature = self.get_parameter("temperature").value
+        self._top_p = self.get_parameter("top_p").value
+        self._soul_path = self.get_parameter("soul_path").value
+        self._db_path = self.get_parameter("context_db_path").value
+        self._save_interval = self.get_parameter("save_interval_s").value
+        self._stream = self.get_parameter("stream").value
+
+        # ── Publishers / Subscribers ─────────────────────────────────────────
+        qos = QoSProfile(depth=10)
+        self._resp_pub = self.create_publisher(
+            ConversationResponse, "/social/conversation/response", qos
+        )
+        self._transcript_sub = self.create_subscription(
+            SpeechTranscript, "/social/speech/transcript",
+            self._on_transcript, qos
+        )
+
+        # ── State ────────────────────────────────────────────────────────────
+        self._llm = None
+        self._system_prompt = load_system_prompt(self._soul_path)
+        self._ctx_store = ContextStore(self._db_path)
+        self._lock = threading.Lock()
+        self._turn_counter = 0
+        self._generating = False
+        self._last_save = time.time()
+
+        # ── Load LLM in background ────────────────────────────────────────────
+        threading.Thread(target=self._load_llm, daemon=True).start()
+
+        # ── Periodic context save ────────────────────────────────────────────
+        self._save_timer = self.create_timer(self._save_interval, self._save_context)
+
+        self.get_logger().info(
+            f"ConversationNode init (model={self._model_path}, "
+            f"gpu_layers={self._n_gpu}, ctx={self._n_ctx})"
+        )
+
+    # ── Model loading ─────────────────────────────────────────────────────────
+
+    def _load_llm(self) -> None:
+        t0 = time.time()
+        self.get_logger().info(f"Loading LLM: {self._model_path}")
+        try:
+            from llama_cpp import Llama
+            self._llm = Llama(
+                model_path=self._model_path,
+                n_ctx=self._n_ctx,
+                n_gpu_layers=self._n_gpu,
+                n_threads=4,
+                verbose=False,
+            )
+            self.get_logger().info(
+                f"LLM ready ({time.time()-t0:.1f}s). "
+                f"Context: {self._n_ctx} tokens, GPU layers: {self._n_gpu}"
+            )
+        except Exception as e:
+            self.get_logger().error(f"LLM load failed: {e}")
+
+    # ── Transcript callback ───────────────────────────────────────────────────
+
+    def _on_transcript(self, msg: SpeechTranscript) -> None:
+        """Handle final transcripts only (skip streaming partials)."""
+        if msg.is_partial:
+            return
+        if not msg.text.strip():
+            return
+
+        self.get_logger().info(
+            f"Transcript [{msg.speaker_id}]: '{msg.text}'"
+        )
+
+        threading.Thread(
+            target=self._generate_response,
+            args=(msg.text.strip(), msg.speaker_id),
+            daemon=True,
+        ).start()
+
+    # ── LLM inference ─────────────────────────────────────────────────────────
+
+    def _generate_response(self, user_text: str, speaker_id: str) -> None:
+        """Generate LLM response with streaming. Runs in thread."""
+        if self._llm is None:
+            self.get_logger().warn("LLM not loaded yet, dropping utterance")
+            return
+
+        with self._lock:
+            if self._generating:
+                self.get_logger().warn("LLM busy, dropping utterance")
+                return
+            self._generating = True
+            self._turn_counter += 1
+            turn_id = self._turn_counter
+
+        try:
+            ctx = self._ctx_store.get(speaker_id)
+
+            # Summary compression if context is long
+            if ctx.needs_compression():
+                self._compress_context(ctx)
+
+            ctx.add_user(user_text)
+
+            prompt = build_llama_prompt(
+                ctx, user_text, self._system_prompt
+            )
+
+            t0 = time.perf_counter()
+            full_response = ""
+
+            if self._stream:
+                output = self._llm(
+                    prompt,
+                    max_tokens=self._max_tokens,
+                    temperature=self._temperature,
+                    top_p=self._top_p,
+                    stream=True,
+                    stop=["<|user|>", "<|system|>", "\n\n\n"],
+                )
+                for chunk in output:
+                    token = chunk["choices"][0]["text"]
+                    full_response += token
+                    # Publish partial after each sentence boundary for low TTS latency
+                    if token.endswith((".", "!", "?", "\n")):
+                        self._publish_response(
+                            full_response.strip(), speaker_id, turn_id, is_partial=True
+                        )
+            else:
+                output = self._llm(
+                    prompt,
+                    max_tokens=self._max_tokens,
+                    temperature=self._temperature,
+                    top_p=self._top_p,
+                    stream=False,
+                    stop=["<|user|>", "<|system|>"],
+                )
+                full_response = output["choices"][0]["text"]
+
+            full_response = full_response.strip()
+            latency_ms = (time.perf_counter() - t0) * 1000
+            self.get_logger().info(
+                f"LLM [{speaker_id}] ({latency_ms:.0f}ms): '{full_response[:80]}'"
+            )
+
+            ctx.add_assistant(full_response)
+            self._publish_response(full_response, speaker_id, turn_id, is_partial=False)
+
+        except Exception as e:
+            self.get_logger().error(f"LLM inference error: {e}")
+        finally:
+            with self._lock:
+                self._generating = False
+
+    def _compress_context(self, ctx) -> None:
+        """Ask LLM to summarize old turns for context compression."""
+        if self._llm is None:
+            ctx.compress("(history omitted)")
+            return
+        try:
+            summary_prompt = needs_summary_prompt(ctx)
+            result = self._llm(summary_prompt, max_tokens=80, temperature=0.3, stream=False)
+            summary = result["choices"][0]["text"].strip()
+            ctx.compress(summary)
+            self.get_logger().debug(
+                f"Context compressed for {ctx.person_id}: '{summary[:60]}'"
+            )
+        except Exception:
+            ctx.compress("(history omitted)")
+
+    # ── Publish ───────────────────────────────────────────────────────────────
+
+    def _publish_response(
+        self, text: str, speaker_id: str, turn_id: int, is_partial: bool
+    ) -> None:
+        msg = ConversationResponse()
+        msg.header.stamp = self.get_clock().now().to_msg()
+        msg.text = text
+        msg.speaker_id = speaker_id
+        msg.is_partial = is_partial
+        msg.turn_id = turn_id
+        self._resp_pub.publish(msg)
+
+    def _save_context(self) -> None:
+        try:
+            self._ctx_store.save()
+        except Exception as e:
+            self.get_logger().error(f"Context save error: {e}")
+
+    def destroy_node(self) -> None:
+        self._save_context()
+        super().destroy_node()
+
+
+def main(args=None) -> None:
+    rclpy.init(args=args)
+    node = ConversationNode()
+    try:
+        rclpy.spin(node)
+    except KeyboardInterrupt:
+        pass
+    finally:
+        node.destroy_node()
+        rclpy.shutdown()

jetson/ros2_ws/src/saltybot_social/saltybot_social/llm_context.py

@@ -0,0 +1,176 @@
+"""llm_context.py — Per-person conversation context management.
+
+No ROS2 dependencies. Manages sliding-window conversation history per person_id,
+builds llama.cpp prompt, handles summary compression when context is full.
+
+Used by conversation_node.py.
+Tested by test/test_llm_context.py.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import time
+from typing import List, Optional, Tuple
+
+# ── Constants ──────────────────────────────────────────────────────────────────
+MAX_TURNS = 20          # Max turns to keep before summary compression
+SUMMARY_KEEP = 4        # Keep last N turns after summarizing older ones
+SYSTEM_PROMPT_PATH = "/soul/SOUL.md"
+DEFAULT_SYSTEM_PROMPT = (
+    "You are Salty, a friendly social robot. "
+    "You are helpful, warm, and concise. "
+    "Keep responses under 2 sentences unless asked to elaborate."
+)
+
+
+# ── Conversation history ───────────────────────────────────────────────────────
+
+class PersonContext:
+    """Conversation history + relationship memory for one person."""
+
+    def __init__(self, person_id: str, person_name: str = "") -> None:
+        self.person_id = person_id
+        self.person_name = person_name or person_id
+        self.turns: List[dict] = []  # {"role": "user"|"assistant", "content": str, "ts": float}
+        self.summary: str = ""
+        self.last_seen: float = time.time()
+        self.interaction_count: int = 0
+
+    def add_user(self, text: str) -> None:
+        self.turns.append({"role": "user", "content": text, "ts": time.time()})
+        self.last_seen = time.time()
+        self.interaction_count += 1
+
+    def add_assistant(self, text: str) -> None:
+        self.turns.append({"role": "assistant", "content": text, "ts": time.time()})
+
+    def needs_compression(self) -> bool:
+        return len(self.turns) > MAX_TURNS
+
+    def compress(self, summary_text: str) -> None:
+        """Replace old turns with a summary, keeping last SUMMARY_KEEP turns."""
+        old_summary = self.summary
+        if old_summary:
+            self.summary = old_summary + " " + summary_text
+        else:
+            self.summary = summary_text
+        # Keep most recent turns
+        self.turns = self.turns[-SUMMARY_KEEP:]
+
+    def to_dict(self) -> dict:
+        return {
+            "person_id": self.person_id,
+            "person_name": self.person_name,
+            "turns": self.turns,
+            "summary": self.summary,
+            "last_seen": self.last_seen,
+            "interaction_count": self.interaction_count,
+        }
+
+    @classmethod
+    def from_dict(cls, d: dict) -> "PersonContext":
+        ctx = cls(d["person_id"], d.get("person_name", ""))
+        ctx.turns = d.get("turns", [])
+        ctx.summary = d.get("summary", "")
+        ctx.last_seen = d.get("last_seen", 0.0)
+        ctx.interaction_count = d.get("interaction_count", 0)
+        return ctx
+
+
+# ── Context store ──────────────────────────────────────────────────────────────
+
+class ContextStore:
+    """Persistent per-person conversation context with optional JSON backing."""
+
+    def __init__(self, db_path: str = "/social_db/conversation_context.json") -> None:
+        self._db_path = db_path
+        self._contexts: dict = {}   # person_id → PersonContext
+        self._load()
+
+    def get(self, person_id: str, person_name: str = "") -> PersonContext:
+        if person_id not in self._contexts:
+            self._contexts[person_id] = PersonContext(person_id, person_name)
+        elif person_name and not self._contexts[person_id].person_name:
+            self._contexts[person_id].person_name = person_name
+        return self._contexts[person_id]
+
+    def save(self) -> None:
+        os.makedirs(os.path.dirname(self._db_path), exist_ok=True)
+        with open(self._db_path, "w") as f:
+            json.dump(
+                {pid: ctx.to_dict() for pid, ctx in self._contexts.items()},
+                f, indent=2
+            )
+
+    def _load(self) -> None:
+        if os.path.exists(self._db_path):
+            try:
+                with open(self._db_path) as f:
+                    raw = json.load(f)
+                self._contexts = {
+                    pid: PersonContext.from_dict(d) for pid, d in raw.items()
+                }
+            except Exception:
+                self._contexts = {}
+
+    def all_persons(self) -> List[str]:
+        return list(self._contexts.keys())
+
+
+# ── Prompt builder ─────────────────────────────────────────────────────────────
+
+def load_system_prompt(path: str = SYSTEM_PROMPT_PATH) -> str:
+    """Load SOUL.md as system prompt. Falls back to default."""
+    if os.path.exists(path):
+        with open(path) as f:
+            return f.read().strip()
+    return DEFAULT_SYSTEM_PROMPT
+
+
+def build_llama_prompt(
+    ctx: PersonContext,
+    user_text: str,
+    system_prompt: str,
+    group_persons: Optional[List[str]] = None,
+) -> str:
+    """Build Phi-3 / Llama-3 style chat prompt string.
+
+    Uses the ChatML format supported by llama-cpp-python.
+    """
+    lines = [f"<|system|>\n{system_prompt}"]
+
+    if ctx.summary:
+        lines.append(
+            f"<|system|>\n[Memory of {ctx.person_name}]: {ctx.summary}"
+        )
+
+    if group_persons:
+        others = [p for p in group_persons if p != ctx.person_id]
+        if others:
+            lines.append(
+                f"<|system|>\n[Also present]: {', '.join(others)}"
+            )
+
+    for turn in ctx.turns:
+        role = "user" if turn["role"] == "user" else "assistant"
+        lines.append(f"<|{role}|>\n{turn['content']}")
+
+    lines.append(f"<|user|>\n{user_text}")
+    lines.append("<|assistant|>")
+
+    return "\n".join(lines)
+
+
+def needs_summary_prompt(ctx: PersonContext) -> str:
+    """Build a prompt asking the LLM to summarize old conversation turns."""
+    turns_text = "\n".join(
+        f"{t['role'].upper()}: {t['content']}"
+        for t in ctx.turns[:-SUMMARY_KEEP]
+    )
+    return (
+        f"<|system|>\nSummarize this conversation between {ctx.person_name} and "
+        "Salty in one sentence, focusing on what was discussed and their relationship.\n"
+        f"<|user|>\n{turns_text}\n<|assistant|>"
+    )

jetson/ros2_ws/src/saltybot_social/saltybot_social/orchestrator_node.py

@@ -0,0 +1,298 @@
+"""orchestrator_node.py — End-to-end social-bot pipeline orchestrator.
+
+Issue #89: Main loop, GPU memory scheduling, latency profiling, watchdog.
+
+State machine:
+  IDLE → LISTENING (wake word) → THINKING (LLM generating) → SPEAKING (TTS)
+       ↑___________________________________________|
+
+Responsibilities:
+  - Owns the top-level pipeline state machine
+  - Monitors GPU memory and throttles inference when approaching limit
+  - Profiles end-to-end latency per interaction turn
+  - Health watchdog: restart speech pipeline on hang
+  - Exposes /social/orchestrator/state for UI / other nodes
+
+ROS2 topics:
+  Subscribe: /social/speech/vad_state   (VadState)
+             /social/speech/transcript  (SpeechTranscript)
+             /social/conversation/response (ConversationResponse)
+  Publish:   /social/orchestrator/state  (std_msgs/String) — JSON status blob
+
+Parameters:
+  gpu_mem_warn_mb   (float, 4000.0) — warn at this free GPU memory (MB)
+  gpu_mem_throttle_mb (float, 2000.0) — suspend new inference requests
+  watchdog_timeout_s  (float, 30.0) — restart speech node if no VAD in N seconds
+  latency_window      (int, 20)      — rolling window for latency stats
+  profile_enabled     (bool, true)
+"""
+
+from __future__ import annotations
+
+import json
+import threading
+import time
+from collections import deque
+from enum import Enum
+from typing import Optional
+
+import rclpy
+from rclpy.node import Node
+from rclpy.qos import QoSProfile
+from std_msgs.msg import String
+
+from saltybot_social_msgs.msg import VadState, SpeechTranscript, ConversationResponse
+
+
+# ── State machine ──────────────────────────────────────────────────────────────
+
+class PipelineState(Enum):
+    IDLE = "idle"
+    LISTENING = "listening"
+    THINKING = "thinking"
+    SPEAKING = "speaking"
+    THROTTLED = "throttled"    # GPU OOM risk — waiting for memory
+
+
+# ── Latency tracker ────────────────────────────────────────────────────────────
+
+class LatencyTracker:
+    """Rolling window latency statistics per pipeline stage."""
+
+    def __init__(self, window: int = 20) -> None:
+        self._window = window
+        self._wakeword_to_transcript: deque = deque(maxlen=window)
+        self._transcript_to_llm_first: deque = deque(maxlen=window)
+        self._llm_first_to_tts_first: deque = deque(maxlen=window)
+        self._end_to_end: deque = deque(maxlen=window)
+
+    def record(self, stage: str, latency_ms: float) -> None:
+        mapping = {
+            "wakeword_to_transcript": self._wakeword_to_transcript,
+            "transcript_to_llm": self._transcript_to_llm_first,
+            "llm_to_tts": self._llm_first_to_tts_first,
+            "end_to_end": self._end_to_end,
+        }
+        q = mapping.get(stage)
+        if q is not None:
+            q.append(latency_ms)
+
+    def stats(self) -> dict:
+        def _stat(q: deque) -> dict:
+            if not q:
+                return {}
+            s = sorted(q)
+            return {
+                "mean_ms": round(sum(s) / len(s), 1),
+                "p50_ms": round(s[len(s) // 2], 1),
+                "p95_ms": round(s[int(len(s) * 0.95)], 1),
+                "n": len(s),
+            }
+        return {
+            "wakeword_to_transcript": _stat(self._wakeword_to_transcript),
+            "transcript_to_llm": _stat(self._transcript_to_llm_first),
+            "llm_to_tts": _stat(self._llm_first_to_tts_first),
+            "end_to_end": _stat(self._end_to_end),
+        }
+
+
+# ── GPU memory helpers ────────────────────────────────────────────────────────
+
+def get_gpu_free_mb() -> Optional[float]:
+    """Return free GPU memory in MB, or None if unavailable."""
+    try:
+        import pycuda.driver as drv
+        drv.init()
+        ctx = drv.Device(0).make_context()
+        free, _ = drv.mem_get_info()
+        ctx.pop()
+        return free / 1024 / 1024
+    except Exception:
+        pass
+    try:
+        import subprocess
+        out = subprocess.check_output(
+            ["nvidia-smi", "--query-gpu=memory.free", "--format=csv,noheader,nounits"],
+            timeout=2
+        ).decode().strip()
+        return float(out.split("\n")[0])
+    except Exception:
+        return None
+
+
+# ── Orchestrator ───────────────────────────────────────────────────────────────
+
+class OrchestratorNode(Node):
+    """Pipeline orchestrator — state machine, GPU watchdog, latency profiler."""
+
+    def __init__(self) -> None:
+        super().__init__("orchestrator_node")
+
+        # ── Parameters ──────────────────────────────────────────────────────
+        self.declare_parameter("gpu_mem_warn_mb", 4000.0)
+        self.declare_parameter("gpu_mem_throttle_mb", 2000.0)
+        self.declare_parameter("watchdog_timeout_s", 30.0)
+        self.declare_parameter("latency_window", 20)
+        self.declare_parameter("profile_enabled", True)
+        self.declare_parameter("state_publish_rate", 2.0)   # Hz
+
+        self._gpu_warn = self.get_parameter("gpu_mem_warn_mb").value
+        self._gpu_throttle = self.get_parameter("gpu_mem_throttle_mb").value
+        self._watchdog_timeout = self.get_parameter("watchdog_timeout_s").value
+        self._profile = self.get_parameter("profile_enabled").value
+        self._state_rate = self.get_parameter("state_publish_rate").value
+
+        # ── Publishers / Subscribers ─────────────────────────────────────────
+        qos = QoSProfile(depth=10)
+        self._state_pub = self.create_publisher(String, "/social/orchestrator/state", qos)
+
+        self._vad_sub = self.create_subscription(
+            VadState, "/social/speech/vad_state", self._on_vad, qos
+        )
+        self._transcript_sub = self.create_subscription(
+            SpeechTranscript, "/social/speech/transcript", self._on_transcript, qos
+        )
+        self._response_sub = self.create_subscription(
+            ConversationResponse, "/social/conversation/response", self._on_response, qos
+        )
+
+        # ── State ────────────────────────────────────────────────────────────
+        self._state = PipelineState.IDLE
+        self._state_lock = threading.Lock()
+        self._tracker = LatencyTracker(self.get_parameter("latency_window").value)
+
+        # Timestamps for latency tracking
+        self._t_wake: float = 0.0
+        self._t_transcript: float = 0.0
+        self._t_llm_first: float = 0.0
+        self._t_tts_first: float = 0.0
+
+        self._last_vad_t: float = time.time()
+        self._last_transcript_t: float = 0.0
+        self._last_response_t: float = 0.0
+
+        # ── Timers ────────────────────────────────────────────────────────────
+        self._state_timer = self.create_timer(1.0 / self._state_rate, self._publish_state)
+        self._gpu_timer = self.create_timer(5.0, self._check_gpu)
+        self._watchdog_timer = self.create_timer(10.0, self._watchdog)
+
+        self.get_logger().info("OrchestratorNode ready")
+
+    # ── State transitions ─────────────────────────────────────────────────────
+
+    def _set_state(self, new_state: PipelineState) -> None:
+        with self._state_lock:
+            if self._state != new_state:
+                self.get_logger().info(
+                    f"Pipeline: {self._state.value} → {new_state.value}"
+                )
+                self._state = new_state
+
+    # ── Subscribers ───────────────────────────────────────────────────────────
+
+    def _on_vad(self, msg: VadState) -> None:
+        self._last_vad_t = time.time()
+        if msg.state == "wake_word":
+            self._t_wake = time.time()
+            self._set_state(PipelineState.LISTENING)
+        elif msg.speech_active and self._state == PipelineState.IDLE:
+            self._set_state(PipelineState.LISTENING)
+
+    def _on_transcript(self, msg: SpeechTranscript) -> None:
+        if msg.is_partial:
+            return
+        self._last_transcript_t = time.time()
+        self._t_transcript = time.time()
+        self._set_state(PipelineState.THINKING)
+
+        if self._profile and self._t_wake > 0:
+            latency_ms = (self._t_transcript - self._t_wake) * 1000
+            self._tracker.record("wakeword_to_transcript", latency_ms)
+            if latency_ms > 500:
+                self.get_logger().warn(
+                    f"Wake→transcript latency high: {latency_ms:.0f}ms (target <500ms)"
+                )
+
+    def _on_response(self, msg: ConversationResponse) -> None:
+        self._last_response_t = time.time()
+
+        if msg.is_partial and self._t_llm_first == 0.0:
+            # First token from LLM
+            self._t_llm_first = time.time()
+            if self._profile and self._t_transcript > 0:
+                latency_ms = (self._t_llm_first - self._t_transcript) * 1000
+                self._tracker.record("transcript_to_llm", latency_ms)
+
+            self._set_state(PipelineState.SPEAKING)
+
+        if not msg.is_partial:
+            # Full response complete
+            self._t_llm_first = 0.0
+            self._t_wake = 0.0
+            self._t_transcript = 0.0
+
+            if self._profile and self._t_wake > 0:
+                e2e_ms = (time.time() - self._t_wake) * 1000
+                self._tracker.record("end_to_end", e2e_ms)
+
+            # Return to IDLE after a short delay (TTS still playing)
+            threading.Timer(2.0, lambda: self._set_state(PipelineState.IDLE)).start()
+
+    # ── GPU memory check ──────────────────────────────────────────────────────
+
+    def _check_gpu(self) -> None:
+        free_mb = get_gpu_free_mb()
+        if free_mb is None:
+            return
+
+        if free_mb < self._gpu_throttle:
+            self.get_logger().error(
+                f"GPU memory critical: {free_mb:.0f}MB free < {self._gpu_throttle:.0f}MB threshold"
+            )
+            self._set_state(PipelineState.THROTTLED)
+        elif free_mb < self._gpu_warn:
+            self.get_logger().warn(f"GPU memory low: {free_mb:.0f}MB free")
+        else:
+            # Recover from throttled state
+            with self._state_lock:
+                if self._state == PipelineState.THROTTLED:
+                    self._set_state(PipelineState.IDLE)
+
+    # ── Watchdog ──────────────────────────────────────────────────────────────
+
+    def _watchdog(self) -> None:
+        """Alert if speech pipeline has gone silent for too long."""
+        since_vad = time.time() - self._last_vad_t
+        if since_vad > self._watchdog_timeout:
+            self.get_logger().error(
+                f"Watchdog: No VAD signal for {since_vad:.0f}s. "
+                "Speech pipeline may be hung. Check /social/speech/vad_state."
+            )
+
+    # ── State publisher ───────────────────────────────────────────────────────
+
+    def _publish_state(self) -> None:
+        with self._state_lock:
+            current = self._state.value
+
+        payload = {
+            "state": current,
+            "ts": time.time(),
+            "latency": self._tracker.stats() if self._profile else {},
+            "gpu_free_mb": get_gpu_free_mb(),
+        }
+        msg = String()
+        msg.data = json.dumps(payload)
+        self._state_pub.publish(msg)
+
+
+def main(args=None) -> None:
+    rclpy.init(args=args)
+    node = OrchestratorNode()
+    try:
+        rclpy.spin(node)
+    except KeyboardInterrupt:
+        pass
+    finally:
+        node.destroy_node()
+        rclpy.shutdown()

jetson/ros2_ws/src/saltybot_social/saltybot_social/speech_pipeline_node.py

@@ -0,0 +1,391 @@
+"""speech_pipeline_node.py — Wake word + VAD + Whisper STT + speaker diarization.
+
+Issue #81: Speech pipeline for social-bot.
+
+Pipeline:
+  USB mic (PyAudio) → VAD (Silero / energy fallback)
+    → wake word gate (OpenWakeWord "hey_salty")
+      → utterance buffer → faster-whisper STT (Orin GPU)
+        → ECAPA-TDNN speaker embedding → identify_speaker()
+          → publish /social/speech/transcript + /social/speech/vad_state
+
+Latency target: <500ms wake-to-first-token (streaming partial transcripts).
+
+ROS2 topics published:
+  /social/speech/transcript  (saltybot_social_msgs/SpeechTranscript)
+  /social/speech/vad_state   (saltybot_social_msgs/VadState)
+
+Parameters:
+  mic_device_index  (int, -1 = system default)
+  sample_rate       (int, 16000)
+  wake_word_model   (str, "hey_salty" — OpenWakeWord model name or path)
+  wake_word_threshold (float, 0.5)
+  vad_threshold_db  (float, -35.0) — fallback energy VAD
+  use_silero_vad    (bool, true)
+  whisper_model     (str, "small")
+  whisper_compute_type (str, "float16")
+  speaker_threshold (float, 0.65) — cosine similarity threshold
+  speaker_db_path   (str, "/social_db/speaker_embeddings.json")
+  publish_partial   (bool, true)
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import threading
+import time
+from typing import Optional
+
+import rclpy
+from rclpy.node import Node
+from rclpy.qos import QoSProfile, QoSReliabilityPolicy, QoSDurabilityPolicy
+from builtin_interfaces.msg import Time
+from std_msgs.msg import Header
+
+from saltybot_social_msgs.msg import SpeechTranscript, VadState
+from .speech_utils import (
+    EnergyVad, UtteranceSegmenter,
+    pcm16_to_float32, rms_db, identify_speaker,
+    SAMPLE_RATE, CHUNK_SAMPLES,
+)
+
+
+class SpeechPipelineNode(Node):
+    """Wake word → VAD → STT → diarization → ROS2 publisher."""
+
+    def __init__(self) -> None:
+        super().__init__("speech_pipeline_node")
+
+        # ── Parameters ──────────────────────────────────────────────────────
+        self.declare_parameter("mic_device_index", -1)
+        self.declare_parameter("sample_rate", SAMPLE_RATE)
+        self.declare_parameter("wake_word_model", "hey_salty")
+        self.declare_parameter("wake_word_threshold", 0.5)
+        self.declare_parameter("vad_threshold_db", -35.0)
+        self.declare_parameter("use_silero_vad", True)
+        self.declare_parameter("whisper_model", "small")
+        self.declare_parameter("whisper_compute_type", "float16")
+        self.declare_parameter("speaker_threshold", 0.65)
+        self.declare_parameter("speaker_db_path", "/social_db/speaker_embeddings.json")
+        self.declare_parameter("publish_partial", True)
+
+        self._mic_idx = self.get_parameter("mic_device_index").value
+        self._rate = self.get_parameter("sample_rate").value
+        self._ww_model = self.get_parameter("wake_word_model").value
+        self._ww_thresh = self.get_parameter("wake_word_threshold").value
+        self._vad_thresh_db = self.get_parameter("vad_threshold_db").value
+        self._use_silero = self.get_parameter("use_silero_vad").value
+        self._whisper_model_name = self.get_parameter("whisper_model").value
+        self._compute_type = self.get_parameter("whisper_compute_type").value
+        self._speaker_thresh = self.get_parameter("speaker_threshold").value
+        self._speaker_db = self.get_parameter("speaker_db_path").value
+        self._publish_partial = self.get_parameter("publish_partial").value
+
+        # ── Publishers ───────────────────────────────────────────────────────
+        qos = QoSProfile(depth=10)
+        self._transcript_pub = self.create_publisher(SpeechTranscript,
+                                                     "/social/speech/transcript", qos)
+        self._vad_pub = self.create_publisher(VadState, "/social/speech/vad_state", qos)
+
+        # ── Internal state ───────────────────────────────────────────────────
+        self._whisper = None
+        self._ecapa = None
+        self._oww = None
+        self._vad = None
+        self._segmenter = None
+        self._known_speakers: dict = {}
+        self._audio_stream = None
+        self._pa = None
+        self._lock = threading.Lock()
+        self._running = False
+        self._wake_word_active = False
+        self._wake_expiry = 0.0
+        self._wake_window_s = 8.0   # seconds to listen after wake word
+        self._turn_counter = 0
+
+        # ── Lazy model loading in background thread ──────────────────────────
+        self._model_thread = threading.Thread(target=self._load_models, daemon=True)
+        self._model_thread.start()
+
+    # ── Model loading ─────────────────────────────────────────────────────────
+
+    def _load_models(self) -> None:
+        """Load all AI models (runs in background to not block ROS spin)."""
+        self.get_logger().info("Loading speech models (Whisper, ECAPA-TDNN, VAD)...")
+        t0 = time.time()
+
+        # faster-whisper
+        try:
+            from faster_whisper import WhisperModel
+            self._whisper = WhisperModel(
+                self._whisper_model_name,
+                device="cuda",
+                compute_type=self._compute_type,
+                download_root="/models",
+            )
+            self.get_logger().info(
+                f"Whisper '{self._whisper_model_name}' loaded "
+                f"({time.time()-t0:.1f}s)"
+            )
+        except Exception as e:
+            self.get_logger().error(f"Whisper load failed: {e}")
+
+        # Silero VAD
+        if self._use_silero:
+            try:
+                from silero_vad import load_silero_vad
+                self._vad = load_silero_vad()
+                self.get_logger().info("Silero VAD loaded")
+            except Exception as e:
+                self.get_logger().warn(f"Silero VAD unavailable ({e}), using energy VAD")
+                self._vad = None
+
+        energy_vad = EnergyVad(threshold_db=self._vad_thresh_db)
+        self._segmenter = UtteranceSegmenter(energy_vad)
+
+        # ECAPA-TDNN speaker embeddings
+        try:
+            from speechbrain.pretrained import EncoderClassifier
+            self._ecapa = EncoderClassifier.from_hparams(
+                source="speechbrain/spkrec-ecapa-voxceleb",
+                savedir="/models/speechbrain_ecapa",
+            )
+            self.get_logger().info("ECAPA-TDNN loaded")
+        except Exception as e:
+            self.get_logger().warn(f"ECAPA-TDNN unavailable ({e}), using 'unknown' speaker")
+
+        # OpenWakeWord
+        try:
+            import openwakeword
+            from openwakeword.model import Model as OWWModel
+            self._oww = OWWModel(
+                wakeword_models=[self._ww_model],
+                inference_framework="onnx",
+            )
+            self.get_logger().info(f"OpenWakeWord '{self._ww_model}' loaded")
+        except Exception as e:
+            self.get_logger().warn(
+                f"OpenWakeWord unavailable ({e}). Robot will listen continuously."
+            )
+
+        # Load speaker DB
+        self._load_speaker_db()
+
+        # Start audio capture
+        self._start_audio()
+        self.get_logger().info(f"Speech pipeline ready ({time.time()-t0:.1f}s total)")
+
+    def _load_speaker_db(self) -> None:
+        """Load speaker embedding database from JSON."""
+        if os.path.exists(self._speaker_db):
+            try:
+                with open(self._speaker_db) as f:
+                    self._known_speakers = json.load(f)
+                self.get_logger().info(
+                    f"Speaker DB loaded: {len(self._known_speakers)} persons"
+                )
+            except Exception as e:
+                self.get_logger().error(f"Speaker DB load error: {e}")
+
+    # ── Audio capture ─────────────────────────────────────────────────────────
+
+    def _start_audio(self) -> None:
+        """Open PyAudio stream and start capture thread."""
+        try:
+            import pyaudio
+            self._pa = pyaudio.PyAudio()
+            dev_idx = None if self._mic_idx < 0 else self._mic_idx
+            self._audio_stream = self._pa.open(
+                format=pyaudio.paInt16,
+                channels=1,
+                rate=self._rate,
+                input=True,
+                input_device_index=dev_idx,
+                frames_per_buffer=CHUNK_SAMPLES,
+            )
+            self._running = True
+            t = threading.Thread(target=self._audio_loop, daemon=True)
+            t.start()
+            self.get_logger().info(
+                f"Audio capture started (device={self._mic_idx}, {self._rate}Hz)"
+            )
+        except Exception as e:
+            self.get_logger().error(f"Audio stream error: {e}")
+
+    def _audio_loop(self) -> None:
+        """Continuous audio capture loop — runs in dedicated thread."""
+        while self._running:
+            try:
+                raw = self._audio_stream.read(CHUNK_SAMPLES, exception_on_overflow=False)
+            except Exception:
+                continue
+
+            samples = pcm16_to_float32(raw)
+            db = rms_db(samples)
+            now = time.time()
+
+            # VAD state publish (10Hz — every 3 chunks at 30ms each)
+            vad_active = self._check_vad(samples)
+            state_str = "silence"
+
+            # Wake word check
+            if self._oww is not None:
+                preds = self._oww.predict(samples)
+                score = preds.get(self._ww_model, 0.0)
+                if isinstance(score, (list, tuple)):
+                    score = score[-1]
+                if score >= self._ww_thresh:
+                    self._wake_word_active = True
+                    self._wake_expiry = now + self._wake_window_s
+                    state_str = "wake_word"
+                    self.get_logger().info(
+                        f"Wake word detected (score={score:.2f})"
+                    )
+            else:
+                # No wake word detector — always listen
+                self._wake_word_active = True
+                self._wake_expiry = now + self._wake_window_s
+
+            if self._wake_word_active and now > self._wake_expiry:
+                self._wake_word_active = False
+
+            # Publish VAD state
+            if vad_active:
+                state_str = "speech"
+            self._publish_vad(vad_active, db, state_str)
+
+            # Only feed segmenter when wake word is active
+            if not self._wake_word_active:
+                continue
+
+            if self._segmenter is None:
+                continue
+
+            completed = self._segmenter.push(samples)
+            for utt_samples, duration in completed:
+                threading.Thread(
+                    target=self._process_utterance,
+                    args=(utt_samples, duration),
+                    daemon=True,
+                ).start()
+
+    def _check_vad(self, samples: list) -> bool:
+        """Run Silero VAD or fall back to segmenter's energy VAD."""
+        if self._vad is not None:
+            try:
+                import torch
+                t = torch.tensor(samples, dtype=torch.float32)
+                prob = self._vad(t, self._rate).item()
+                return prob > 0.5
+            except Exception:
+                pass
+        if self._segmenter is not None:
+            return self._segmenter._vad.is_active
+        return False
+
+    # ── STT + diarization ─────────────────────────────────────────────────────
+
+    def _process_utterance(self, samples: list, duration: float) -> None:
+        """Transcribe utterance and identify speaker. Runs in thread."""
+        if self._whisper is None:
+            return
+
+        t0 = time.perf_counter()
+
+        # Convert to numpy for faster-whisper
+        try:
+            import numpy as np
+            audio_np = np.array(samples, dtype=np.float32)
+        except ImportError:
+            audio_np = samples
+
+        # Speaker embedding first (can be concurrent with transcription)
+        speaker_id = "unknown"
+        if self._ecapa is not None:
+            try:
+                import torch
+                audio_tensor = torch.tensor([samples], dtype=torch.float32)
+                with torch.no_grad():
+                    emb = self._ecapa.encode_batch(audio_tensor)
+                emb_list = emb[0].cpu().numpy().tolist()
+                speaker_id = identify_speaker(
+                    emb_list, self._known_speakers, self._speaker_thresh
+                )
+            except Exception as e:
+                self.get_logger().debug(f"Speaker ID error: {e}")
+
+        # Streaming Whisper transcription
+        partial_text = ""
+        try:
+            segments_gen, _info = self._whisper.transcribe(
+                audio_np,
+                language="en",
+                beam_size=3,
+                vad_filter=False,
+            )
+            for seg in segments_gen:
+                partial_text += seg.text.strip() + " "
+                if self._publish_partial:
+                    self._publish_transcript(
+                        partial_text.strip(), speaker_id, 0.0, duration, is_partial=True
+                    )
+        except Exception as e:
+            self.get_logger().error(f"Whisper error: {e}")
+            return
+
+        final_text = partial_text.strip()
+        if not final_text:
+            return
+
+        latency_ms = (time.perf_counter() - t0) * 1000
+        self.get_logger().info(
+            f"STT [{speaker_id}] ({duration:.1f}s, {latency_ms:.0f}ms): '{final_text}'"
+        )
+        self._publish_transcript(final_text, speaker_id, 0.9, duration, is_partial=False)
+
+    # ── Publishers ────────────────────────────────────────────────────────────
+
+    def _publish_transcript(
+        self, text: str, speaker_id: str, confidence: float,
+        duration: float, is_partial: bool
+    ) -> None:
+        msg = SpeechTranscript()
+        msg.header.stamp = self.get_clock().now().to_msg()
+        msg.text = text
+        msg.speaker_id = speaker_id
+        msg.confidence = confidence
+        msg.audio_duration = duration
+        msg.is_partial = is_partial
+        self._transcript_pub.publish(msg)
+
+    def _publish_vad(self, speech_active: bool, db: float, state: str) -> None:
+        msg = VadState()
+        msg.header.stamp = self.get_clock().now().to_msg()
+        msg.speech_active = speech_active
+        msg.energy_db = db
+        msg.state = state
+        self._vad_pub.publish(msg)
+
+    # ── Cleanup ───────────────────────────────────────────────────────────────
+
+    def destroy_node(self) -> None:
+        self._running = False
+        if self._audio_stream:
+            self._audio_stream.stop_stream()
+            self._audio_stream.close()
+        if self._pa:
+            self._pa.terminate()
+        super().destroy_node()
+
+
+def main(args=None) -> None:
+    rclpy.init(args=args)
+    node = SpeechPipelineNode()
+    try:
+        rclpy.spin(node)
+    except KeyboardInterrupt:
+        pass
+    finally:
+        node.destroy_node()
+        rclpy.shutdown()

jetson/ros2_ws/src/saltybot_social/saltybot_social/speech_utils.py

@@ -0,0 +1,188 @@
+"""speech_utils.py — Pure helpers for wake word, VAD, and STT.
+
+No ROS2 dependencies. All functions operate on raw numpy float32 arrays
+at 16 kHz mono (the standard for Whisper / Silero VAD).
+
+Tested by test/test_speech_utils.py (no GPU required).
+"""
+
+from __future__ import annotations
+
+import math
+import struct
+from typing import Generator, List, Tuple
+
+# ── Constants ──────────────────────────────────────────────────────────────────
+SAMPLE_RATE = 16000          # Hz — standard for Whisper / Silero
+CHUNK_MS = 30                # VAD frame size
+CHUNK_SAMPLES = int(SAMPLE_RATE * CHUNK_MS / 1000)   # 480 samples
+INT16_MAX = 32768.0
+
+
+# ── Audio helpers ──────────────────────────────────────────────────────────────
+
+def pcm16_to_float32(data: bytes) -> list:
+    """Convert raw PCM16 LE bytes → float32 list in [-1.0, 1.0]."""
+    n = len(data) // 2
+    samples = struct.unpack(f"<{n}h", data[:n * 2])
+    return [s / INT16_MAX for s in samples]
+
+
+def float32_to_pcm16(samples: list) -> bytes:
+    """Convert float32 list → PCM16 LE bytes (clipped to ±1.0)."""
+    clipped = [max(-1.0, min(1.0, s)) for s in samples]
+    ints = [max(-32768, min(32767, int(s * INT16_MAX))) for s in clipped]
+    return struct.pack(f"<{len(ints)}h", *ints)
+
+
+def rms_db(samples: list) -> float:
+    """Compute RMS energy in dBFS. Returns -96.0 for silence."""
+    if not samples:
+        return -96.0
+    mean_sq = sum(s * s for s in samples) / len(samples)
+    rms = math.sqrt(mean_sq) if mean_sq > 0.0 else 1e-10
+    return 20.0 * math.log10(max(rms, 1e-10))
+
+
+def chunk_audio(samples: list, chunk_size: int = CHUNK_SAMPLES) -> Generator[list, None, None]:
+    """Yield fixed-size chunks from a flat sample list."""
+    for i in range(0, len(samples) - chunk_size + 1, chunk_size):
+        yield samples[i:i + chunk_size]
+
+
+# ── Simple energy-based VAD (fallback when Silero unavailable) ────────────────
+
+class EnergyVad:
+    """Threshold-based energy VAD.
+
+    Useful as a lightweight fallback / unit-testable reference.
+    """
+
+    def __init__(
+        self,
+        threshold_db: float = -35.0,
+        onset_frames: int = 2,
+        offset_frames: int = 8,
+    ) -> None:
+        self.threshold_db = threshold_db
+        self.onset_frames = onset_frames
+        self.offset_frames = offset_frames
+        self._above_count = 0
+        self._below_count = 0
+        self._active = False
+
+    def process(self, chunk: list) -> bool:
+        """Process one chunk (CHUNK_SAMPLES long). Returns True if speech active."""
+        db = rms_db(chunk)
+        if db >= self.threshold_db:
+            self._above_count += 1
+            self._below_count = 0
+            if self._above_count >= self.onset_frames:
+                self._active = True
+        else:
+            self._below_count += 1
+            self._above_count = 0
+            if self._below_count >= self.offset_frames:
+                self._active = False
+        return self._active
+
+    def reset(self) -> None:
+        self._above_count = 0
+        self._below_count = 0
+        self._active = False
+
+    @property
+    def is_active(self) -> bool:
+        return self._active
+
+
+# ── Utterance segmenter ────────────────────────────────────────────────────────
+
+class UtteranceSegmenter:
+    """Accumulates audio chunks into complete utterances using VAD.
+
+    Yields (samples, duration_s) when an utterance ends.
+    Pre-roll keeps a buffer before speech onset (catches first phoneme).
+    """
+
+    def __init__(
+        self,
+        vad: EnergyVad,
+        pre_roll_frames: int = 5,
+        max_duration_s: float = 15.0,
+        sample_rate: int = SAMPLE_RATE,
+        chunk_samples: int = CHUNK_SAMPLES,
+    ) -> None:
+        self._vad = vad
+        self._pre_roll = pre_roll_frames
+        self._max_frames = int(max_duration_s * sample_rate / chunk_samples)
+        self._sample_rate = sample_rate
+        self._chunk_samples = chunk_samples
+        self._pre_buf: list = []   # ring buffer for pre-roll
+        self._utt_buf: list = []   # growing utterance buffer
+        self._in_utt = False
+        self._silence_after = 0
+
+    def push(self, chunk: list) -> list:
+        """Push one chunk, return list of completed (samples, duration_s) tuples."""
+        speech = self._vad.process(chunk)
+        results = []
+
+        if not self._in_utt:
+            self._pre_buf.append(chunk)
+            if len(self._pre_buf) > self._pre_roll:
+                self._pre_buf.pop(0)
+
+            if speech:
+                self._in_utt = True
+                self._utt_buf = [s for c in self._pre_buf for s in c]
+                self._utt_buf.extend(chunk)
+                self._silence_after = 0
+        else:
+            self._utt_buf.extend(chunk)
+            if not speech:
+                self._silence_after += 1
+                if (self._silence_after >= self._vad.offset_frames or
+                        len(self._utt_buf) >= self._max_frames * self._chunk_samples):
+                    dur = len(self._utt_buf) / self._sample_rate
+                    results.append((list(self._utt_buf), dur))
+                    self._utt_buf = []
+                    self._in_utt = False
+                    self._silence_after = 0
+            else:
+                self._silence_after = 0
+
+        return results
+
+
+# ── Speaker embedding distance ─────────────────────────────────────────────────
+
+def cosine_similarity(a: list, b: list) -> float:
+    """Cosine similarity between two embedding vectors."""
+    if len(a) != len(b) or not a:
+        return 0.0
+    dot = sum(x * y for x, y in zip(a, b))
+    norm_a = math.sqrt(sum(x * x for x in a))
+    norm_b = math.sqrt(sum(x * x for x in b))
+    if norm_a < 1e-10 or norm_b < 1e-10:
+        return 0.0
+    return dot / (norm_a * norm_b)
+
+
+def identify_speaker(
+    embedding: list,
+    known_speakers: dict,
+    threshold: float = 0.65,
+) -> str:
+    """Return speaker_id from known_speakers dict or 'unknown'.
+
+    known_speakers: {speaker_id: embedding_list}
+    """
+    best_id = "unknown"
+    best_sim = threshold
+    for spk_id, spk_emb in known_speakers.items():
+        sim = cosine_similarity(embedding, spk_emb)
+        if sim > best_sim:
+            best_sim = sim
+            best_id = spk_id
+    return best_id

jetson/ros2_ws/src/saltybot_social/saltybot_social/tts_node.py

@@ -0,0 +1,228 @@
+"""tts_node.py — Streaming TTS with Piper / first-chunk streaming.
+
+Issue #85: Streaming TTS — Piper/XTTS integration with first-chunk streaming.
+
+Pipeline:
+  /social/conversation/response (ConversationResponse)
+    → sentence split → Piper ONNX synthesis (sentence by sentence)
+      → PCM16 chunks → USB speaker (sounddevice) + publish /social/tts/audio
+
+First-chunk strategy:
+  - On partial=true ConversationResponse, extract first sentence and synthesize
+    immediately → audio starts before LLM finishes generating
+  - On final=false, synthesize remaining sentences
+
+Latency target: <200ms to first audio chunk.
+
+ROS2 topics:
+  Subscribe: /social/conversation/response (saltybot_social_msgs/ConversationResponse)
+  Publish:   /social/tts/audio (audio_msgs/Audio or std_msgs/UInt8MultiArray fallback)
+
+Parameters:
+  voice_path      (str, "/models/piper/en_US-lessac-medium.onnx")
+  sample_rate     (int, 22050)
+  volume          (float, 1.0)
+  audio_device    (str, "")  — sounddevice device name; "" = system default
+  playback_enabled (bool, true)
+  publish_audio   (bool, false)  — publish PCM to ROS2 topic
+  sentence_streaming (bool, true) — synthesize sentence-by-sentence
+"""
+
+from __future__ import annotations
+
+import queue
+import threading
+import time
+from typing import Optional
+
+import rclpy
+from rclpy.node import Node
+from rclpy.qos import QoSProfile
+from std_msgs.msg import UInt8MultiArray
+
+from saltybot_social_msgs.msg import ConversationResponse
+from .tts_utils import split_sentences, strip_ssml, apply_volume, chunk_pcm, estimate_duration_ms
+
+
+class TtsNode(Node):
+    """Streaming TTS node using Piper ONNX."""
+
+    def __init__(self) -> None:
+        super().__init__("tts_node")
+
+        # ── Parameters ──────────────────────────────────────────────────────
+        self.declare_parameter("voice_path", "/models/piper/en_US-lessac-medium.onnx")
+        self.declare_parameter("sample_rate", 22050)
+        self.declare_parameter("volume", 1.0)
+        self.declare_parameter("audio_device", "")
+        self.declare_parameter("playback_enabled", True)
+        self.declare_parameter("publish_audio", False)
+        self.declare_parameter("sentence_streaming", True)
+
+        self._voice_path = self.get_parameter("voice_path").value
+        self._sample_rate = self.get_parameter("sample_rate").value
+        self._volume = self.get_parameter("volume").value
+        self._audio_device = self.get_parameter("audio_device").value or None
+        self._playback = self.get_parameter("playback_enabled").value
+        self._publish_audio = self.get_parameter("publish_audio").value
+        self._sentence_streaming = self.get_parameter("sentence_streaming").value
+
+        # ── Publishers / Subscribers ─────────────────────────────────────────
+        qos = QoSProfile(depth=10)
+        self._resp_sub = self.create_subscription(
+            ConversationResponse, "/social/conversation/response",
+            self._on_response, qos
+        )
+        if self._publish_audio:
+            self._audio_pub = self.create_publisher(
+                UInt8MultiArray, "/social/tts/audio", qos
+            )
+
+        # ── TTS engine ────────────────────────────────────────────────────────
+        self._voice = None
+        self._playback_queue: queue.Queue = queue.Queue(maxsize=16)
+        self._current_turn = -1
+        self._synthesized_turns: set = set()   # turn_ids already synthesized
+        self._lock = threading.Lock()
+
+        threading.Thread(target=self._load_voice, daemon=True).start()
+        threading.Thread(target=self._playback_worker, daemon=True).start()
+
+        self.get_logger().info(
+            f"TtsNode init (voice={self._voice_path}, "
+            f"streaming={self._sentence_streaming})"
+        )
+
+    # ── Voice loading ─────────────────────────────────────────────────────────
+
+    def _load_voice(self) -> None:
+        t0 = time.time()
+        self.get_logger().info(f"Loading Piper voice: {self._voice_path}")
+        try:
+            from piper import PiperVoice
+            self._voice = PiperVoice.load(self._voice_path)
+            # Warmup synthesis to pre-JIT ONNX graph
+            warmup_text = "Hello."
+            list(self._voice.synthesize_stream_raw(warmup_text))
+            self.get_logger().info(f"Piper voice ready ({time.time()-t0:.1f}s)")
+        except Exception as e:
+            self.get_logger().error(f"Piper voice load failed: {e}")
+
+    # ── Response handler ──────────────────────────────────────────────────────
+
+    def _on_response(self, msg: ConversationResponse) -> None:
+        """Handle streaming LLM response — synthesize sentence by sentence."""
+        if not msg.text.strip():
+            return
+
+        with self._lock:
+            is_new_turn = msg.turn_id != self._current_turn
+            if is_new_turn:
+                self._current_turn = msg.turn_id
+                # Clear old synthesized sentence cache for this new turn
+                self._synthesized_turns = set()
+
+        text = strip_ssml(msg.text)
+
+        if self._sentence_streaming:
+            sentences = split_sentences(text)
+            for sentence in sentences:
+                # Track which sentences we've already queued by content hash
+                key = (msg.turn_id, hash(sentence))
+                with self._lock:
+                    if key in self._synthesized_turns:
+                        continue
+                    self._synthesized_turns.add(key)
+                self._queue_synthesis(sentence)
+        elif not msg.is_partial:
+            # Non-streaming: synthesize full response at end
+            self._queue_synthesis(text)
+
+    def _queue_synthesis(self, text: str) -> None:
+        """Queue a text segment for synthesis in the playback worker."""
+        if not text.strip():
+            return
+        try:
+            self._playback_queue.put_nowait(text.strip())
+        except queue.Full:
+            self.get_logger().warn("TTS playback queue full, dropping segment")
+
+    # ── Playback worker ───────────────────────────────────────────────────────
+
+    def _playback_worker(self) -> None:
+        """Consume synthesis queue: synthesize → play → publish."""
+        while rclpy.ok():
+            try:
+                text = self._playback_queue.get(timeout=0.5)
+            except queue.Empty:
+                continue
+
+            if self._voice is None:
+                self.get_logger().warn("TTS voice not loaded yet")
+                self._playback_queue.task_done()
+                continue
+
+            t0 = time.perf_counter()
+            pcm_data = self._synthesize(text)
+            if pcm_data is None:
+                self._playback_queue.task_done()
+                continue
+
+            synth_ms = (time.perf_counter() - t0) * 1000
+            dur_ms = estimate_duration_ms(pcm_data, self._sample_rate)
+            self.get_logger().debug(
+                f"TTS '{text[:40]}' synth={synth_ms:.0f}ms, dur={dur_ms:.0f}ms"
+            )
+
+            if self._volume != 1.0:
+                pcm_data = apply_volume(pcm_data, self._volume)
+
+            if self._playback:
+                self._play_audio(pcm_data)
+
+            if self._publish_audio:
+                self._publish_pcm(pcm_data)
+
+            self._playback_queue.task_done()
+
+    def _synthesize(self, text: str) -> Optional[bytes]:
+        """Synthesize text to PCM16 bytes using Piper streaming."""
+        if self._voice is None:
+            return None
+        try:
+            chunks = list(self._voice.synthesize_stream_raw(text))
+            return b"".join(chunks)
+        except Exception as e:
+            self.get_logger().error(f"TTS synthesis error: {e}")
+            return None
+
+    def _play_audio(self, pcm_data: bytes) -> None:
+        """Play PCM16 data on USB speaker via sounddevice."""
+        try:
+            import sounddevice as sd
+            import numpy as np
+            samples = np.frombuffer(pcm_data, dtype=np.int16).astype(np.float32) / 32768.0
+            sd.play(samples, samplerate=self._sample_rate, device=self._audio_device,
+                    blocking=True)
+        except Exception as e:
+            self.get_logger().error(f"Audio playback error: {e}")
+
+    def _publish_pcm(self, pcm_data: bytes) -> None:
+        """Publish PCM data as UInt8MultiArray."""
+        if not hasattr(self, "_audio_pub"):
+            return
+        msg = UInt8MultiArray()
+        msg.data = list(pcm_data)
+        self._audio_pub.publish(msg)
+
+
+def main(args=None) -> None:
+    rclpy.init(args=args)
+    node = TtsNode()
+    try:
+        rclpy.spin(node)
+    except KeyboardInterrupt:
+        pass
+    finally:
+        node.destroy_node()
+        rclpy.shutdown()

jetson/ros2_ws/src/saltybot_social/saltybot_social/tts_utils.py

@@ -0,0 +1,100 @@
+"""tts_utils.py — Pure TTS helpers for streaming synthesis.
+
+No ROS2 dependencies. Provides streaming Piper synthesis and audio chunking.
+
+Tested by test/test_tts_utils.py.
+"""
+
+from __future__ import annotations
+
+import re
+import struct
+from typing import Generator, List, Tuple
+
+
+# ── SSML helpers ──────────────────────────────────────────────────────────────
+
+_SSML_STRIP_RE = re.compile(r"<[^>]+>")
+
+
+def strip_ssml(text: str) -> str:
+    """Remove SSML tags, returning plain text."""
+    return _SSML_STRIP_RE.sub("", text).strip()
+
+
+def split_sentences(text: str) -> List[str]:
+    """Split text into sentences for streaming TTS synthesis.
+
+    Returns list of non-empty sentence strings.
+    """
+    # Split on sentence-ending punctuation followed by whitespace or end
+    parts = re.split(r"(?<=[.!?])\s+", text.strip())
+    return [p.strip() for p in parts if p.strip()]
+
+
+def split_first_sentence(text: str) -> Tuple[str, str]:
+    """Return (first_sentence, remainder).
+
+    Used by TTS node to begin synthesis of first sentence immediately
+    while the LLM continues generating the rest.
+    """
+    sentences = split_sentences(text)
+    if not sentences:
+        return "", ""
+    return sentences[0], " ".join(sentences[1:])
+
+
+# ── PCM audio helpers ─────────────────────────────────────────────────────────
+
+def pcm16_to_wav_bytes(pcm_data: bytes, sample_rate: int = 22050) -> bytes:
+    """Wrap raw PCM16 LE mono data with a WAV header."""
+    num_samples = len(pcm_data) // 2
+    num_channels = 1
+    bits_per_sample = 16
+    byte_rate = sample_rate * num_channels * bits_per_sample // 8
+    block_align = num_channels * bits_per_sample // 8
+    data_size = len(pcm_data)
+    header_size = 44
+
+    header = struct.pack(
+        "<4sI4s4sIHHIIHH4sI",
+        b"RIFF",
+        header_size - 8 + data_size,
+        b"WAVE",
+        b"fmt ",
+        16,                 # chunk size
+        1,                  # PCM format
+        num_channels,
+        sample_rate,
+        byte_rate,
+        block_align,
+        bits_per_sample,
+        b"data",
+        data_size,
+    )
+    return header + pcm_data
+
+
+def chunk_pcm(pcm_data: bytes, chunk_ms: int = 200, sample_rate: int = 22050) -> Generator[bytes, None, None]:
+    """Yield PCM data in fixed-size chunks for streaming playback."""
+    chunk_bytes = (sample_rate * 2 * chunk_ms) // 1000  # int16 = 2 bytes/sample
+    chunk_bytes = max(chunk_bytes, 2) & ~1  # ensure even (int16 aligned)
+    for i in range(0, len(pcm_data), chunk_bytes):
+        yield pcm_data[i:i + chunk_bytes]
+
+
+def estimate_duration_ms(pcm_data: bytes, sample_rate: int = 22050) -> float:
+    """Estimate playback duration in ms from PCM byte length."""
+    return len(pcm_data) / (sample_rate * 2) * 1000.0
+
+
+# ── Volume control ────────────────────────────────────────────────────────────
+
+def apply_volume(pcm_data: bytes, gain: float) -> bytes:
+    """Scale PCM16 LE samples by gain factor (0.0–2.0). Clips to ±32767."""
+    if abs(gain - 1.0) < 0.01:
+        return pcm_data
+    n = len(pcm_data) // 2
+    samples = struct.unpack(f"<{n}h", pcm_data[:n * 2])
+    scaled = [max(-32767, min(32767, int(s * gain))) for s in samples]
+    return struct.pack(f"<{n}h", *scaled)

jetson/ros2_ws/src/saltybot_social/setup.py

@@ -21,7 +21,7 @@ setup(
     zip_safe=True,
     maintainer='seb',
     maintainer_email='seb@vayrette.com',
-    description='Social interaction layer — person state tracking, LED expression + attention',
+    description='Social interaction layer — person tracking, speech, LLM, TTS, orchestrator',
     license='MIT',
     tests_require=['pytest'],
     entry_points={
@@ -29,6 +29,10 @@ setup(
             'person_state_tracker = saltybot_social.person_state_tracker_node:main',
             'expression_node = saltybot_social.expression_node:main',
             'attention_node = saltybot_social.attention_node:main',
+            'speech_pipeline_node = saltybot_social.speech_pipeline_node:main',
+            'conversation_node = saltybot_social.conversation_node:main',
+            'tts_node = saltybot_social.tts_node:main',
+            'orchestrator_node = saltybot_social.orchestrator_node:main',
         ],
     },
 )

jetson/ros2_ws/src/saltybot_social/test/test_llm_context.py

@@ -0,0 +1,244 @@
+"""test_llm_context.py — Unit tests for llm_context.py (no GPU / LLM needed)."""
+
+import sys
+import os
+import tempfile
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+from saltybot_social.llm_context import (
+    PersonContext, ContextStore,
+    build_llama_prompt, MAX_TURNS, SUMMARY_KEEP,
+    DEFAULT_SYSTEM_PROMPT,
+)
+import pytest
+
+
+# ── PersonContext ─────────────────────────────────────────────────────────────
+
+class TestPersonContext:
+    def test_create(self):
+        ctx = PersonContext("person_1", "Alice")
+        assert ctx.person_id == "person_1"
+        assert ctx.person_name == "Alice"
+        assert ctx.turns == []
+        assert ctx.interaction_count == 0
+
+    def test_add_user_increments_count(self):
+        ctx = PersonContext("person_1")
+        ctx.add_user("Hello")
+        assert ctx.interaction_count == 1
+        assert ctx.turns[-1]["role"] == "user"
+        assert ctx.turns[-1]["content"] == "Hello"
+
+    def test_add_assistant(self):
+        ctx = PersonContext("person_1")
+        ctx.add_assistant("Hi there!")
+        assert ctx.turns[-1]["role"] == "assistant"
+
+    def test_needs_compression_false(self):
+        ctx = PersonContext("p")
+        for i in range(MAX_TURNS - 1):
+            ctx.add_user(f"msg {i}")
+        assert not ctx.needs_compression()
+
+    def test_needs_compression_true(self):
+        ctx = PersonContext("p")
+        for i in range(MAX_TURNS + 1):
+            ctx.add_user(f"msg {i}")
+        assert ctx.needs_compression()
+
+    def test_compress_keeps_recent_turns(self):
+        ctx = PersonContext("p")
+        for i in range(MAX_TURNS + 5):
+            ctx.add_user(f"msg {i}")
+        ctx.compress("Summary of conversation")
+        assert len(ctx.turns) == SUMMARY_KEEP
+        assert ctx.summary == "Summary of conversation"
+
+    def test_compress_appends_summary(self):
+        ctx = PersonContext("p")
+        ctx.add_user("msg")
+        ctx.compress("First summary")
+        ctx.add_user("msg2")
+        ctx.compress("Second summary")
+        assert "First summary" in ctx.summary
+        assert "Second summary" in ctx.summary
+
+    def test_roundtrip_serialization(self):
+        ctx = PersonContext("person_42", "Bob")
+        ctx.add_user("test")
+        ctx.add_assistant("response")
+        ctx.summary = "Earlier they discussed robots"
+        d = ctx.to_dict()
+        ctx2 = PersonContext.from_dict(d)
+        assert ctx2.person_id == "person_42"
+        assert ctx2.person_name == "Bob"
+        assert len(ctx2.turns) == 2
+        assert ctx2.summary == "Earlier they discussed robots"
+
+
+# ── ContextStore ──────────────────────────────────────────────────────────────
+
+class TestContextStore:
+    def test_get_creates_new(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            db = os.path.join(tmpdir, "ctx.json")
+            store = ContextStore(db)
+            ctx = store.get("p1", "Alice")
+            assert ctx.person_id == "p1"
+            assert ctx.person_name == "Alice"
+
+    def test_get_same_instance(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            db = os.path.join(tmpdir, "ctx.json")
+            store = ContextStore(db)
+            ctx1 = store.get("p1")
+            ctx2 = store.get("p1")
+            assert ctx1 is ctx2
+
+    def test_save_and_load(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            db = os.path.join(tmpdir, "ctx.json")
+            store = ContextStore(db)
+            ctx = store.get("p1", "Alice")
+            ctx.add_user("Hello")
+            store.save()
+
+            store2 = ContextStore(db)
+            ctx2 = store2.get("p1")
+            assert ctx2.person_name == "Alice"
+            assert len(ctx2.turns) == 1
+            assert ctx2.turns[0]["content"] == "Hello"
+
+    def test_all_persons(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            db = os.path.join(tmpdir, "ctx.json")
+            store = ContextStore(db)
+            store.get("p1")
+            store.get("p2")
+            store.get("p3")
+            assert set(store.all_persons()) == {"p1", "p2", "p3"}
+
+
+# ── Prompt builder ────────────────────────────────────────────────────────────
+
+class TestBuildLlamaPrompt:
+    def test_contains_system_prompt(self):
+        ctx = PersonContext("p1")
+        prompt = build_llama_prompt(ctx, "hello", DEFAULT_SYSTEM_PROMPT)
+        assert DEFAULT_SYSTEM_PROMPT in prompt
+
+    def test_contains_user_text(self):
+        ctx = PersonContext("p1")
+        prompt = build_llama_prompt(ctx, "What time is it?", DEFAULT_SYSTEM_PROMPT)
+        assert "What time is it?" in prompt
+
+    def test_ends_with_assistant_tag(self):
+        ctx = PersonContext("p1")
+        prompt = build_llama_prompt(ctx, "hello", DEFAULT_SYSTEM_PROMPT)
+        assert prompt.strip().endswith("<|assistant|>")
+
+    def test_history_included(self):
+        ctx = PersonContext("p1")
+        ctx.add_user("Previous question")
+        ctx.add_assistant("Previous answer")
+        prompt = build_llama_prompt(ctx, "New question", DEFAULT_SYSTEM_PROMPT)
+        assert "Previous question" in prompt
+        assert "Previous answer" in prompt
+
+    def test_summary_included(self):
+        ctx = PersonContext("p1", "Alice")
+        ctx.summary = "Alice visited last week and likes coffee"
+        prompt = build_llama_prompt(ctx, "hello", DEFAULT_SYSTEM_PROMPT)
+        assert "likes coffee" in prompt
+
+    def test_group_persons_included(self):
+        ctx = PersonContext("p1", "Alice")
+        prompt = build_llama_prompt(ctx, "hello", DEFAULT_SYSTEM_PROMPT,
+                                    group_persons=["p1", "p2_Bob", "p3_Charlie"])
+        assert "p2_Bob" in prompt or "p3_Charlie" in prompt
+
+    def test_no_group_persons_not_in_prompt(self):
+        ctx = PersonContext("p1")
+        prompt = build_llama_prompt(ctx, "hello", DEFAULT_SYSTEM_PROMPT)
+        assert "Also present" not in prompt
+
+
+# ── TTS utils (imported here for coverage) ────────────────────────────────────
+
+class TestTtsUtils:
+    def setup_method(self):
+        sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+        from saltybot_social.tts_utils import (
+            strip_ssml, split_sentences, split_first_sentence,
+            pcm16_to_wav_bytes, chunk_pcm, estimate_duration_ms, apply_volume
+        )
+        self.strip_ssml = strip_ssml
+        self.split_sentences = split_sentences
+        self.split_first_sentence = split_first_sentence
+        self.pcm16_to_wav_bytes = pcm16_to_wav_bytes
+        self.chunk_pcm = chunk_pcm
+        self.estimate_duration_ms = estimate_duration_ms
+        self.apply_volume = apply_volume
+
+    def test_strip_ssml(self):
+        assert self.strip_ssml("<speak>Hello <break time='1s'/> world</speak>") == "Hello  world"
+        assert self.strip_ssml("plain text") == "plain text"
+
+    def test_split_sentences_basic(self):
+        text = "Hello world. How are you? I am fine!"
+        sentences = self.split_sentences(text)
+        assert len(sentences) == 3
+
+    def test_split_sentences_single(self):
+        sentences = self.split_sentences("Hello world.")
+        assert len(sentences) == 1
+        assert sentences[0] == "Hello world."
+
+    def test_split_sentences_empty(self):
+        assert self.split_sentences("") == []
+        assert self.split_sentences("   ") == []
+
+    def test_split_first_sentence(self):
+        text = "Hello world. How are you? I am fine."
+        first, rest = self.split_first_sentence(text)
+        assert first == "Hello world."
+        assert "How are you?" in rest
+
+    def test_wav_header(self):
+        pcm = b"\x00\x00" * 1000
+        wav = self.pcm16_to_wav_bytes(pcm, 22050)
+        assert wav[:4] == b"RIFF"
+        assert wav[8:12] == b"WAVE"
+        assert len(wav) == 44 + len(pcm)
+
+    def test_chunk_pcm_sizes(self):
+        pcm = b"\x00\x00" * 2205  # ~100ms at 22050Hz
+        chunks = list(self.chunk_pcm(pcm, 50, 22050))  # 50ms chunks
+        for chunk in chunks:
+            assert len(chunk) % 2 == 0  # int16 aligned
+
+    def test_estimate_duration(self):
+        pcm = b"\x00\x00" * 22050  # 1s at 22050Hz
+        ms = self.estimate_duration_ms(pcm, 22050)
+        assert abs(ms - 1000.0) < 1.0
+
+    def test_apply_volume_unity(self):
+        pcm = b"\x00\x40" * 100  # some non-zero samples
+        result = self.apply_volume(pcm, 1.0)
+        assert result == pcm
+
+    def test_apply_volume_half(self):
+        import struct
+        pcm = struct.pack("<h", 1000)
+        result = self.apply_volume(pcm, 0.5)
+        val = struct.unpack("<h", result)[0]
+        assert abs(val - 500) <= 1
+
+    def test_apply_volume_clips(self):
+        import struct
+        pcm = struct.pack("<h", 30000)
+        result = self.apply_volume(pcm, 2.0)
+        val = struct.unpack("<h", result)[0]
+        assert val == 32767

jetson/ros2_ws/src/saltybot_social/test/test_speech_utils.py

@@ -0,0 +1,237 @@
+"""test_speech_utils.py — Unit tests for speech_utils.py (no ROS2 / GPU needed)."""
+
+import math
+import struct
+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+from saltybot_social.speech_utils import (
+    pcm16_to_float32, float32_to_pcm16,
+    rms_db, chunk_audio,
+    EnergyVad, UtteranceSegmenter,
+    cosine_similarity, identify_speaker,
+    CHUNK_SAMPLES, SAMPLE_RATE,
+)
+import pytest
+
+
+# ── PCM conversion ────────────────────────────────────────────────────────────
+
+class TestPcmConversion:
+    def test_pcm16_to_float32_silence(self):
+        data = b"\x00\x00" * 100
+        samples = pcm16_to_float32(data)
+        assert len(samples) == 100
+        assert all(s == 0.0 for s in samples)
+
+    def test_pcm16_to_float32_max_positive(self):
+        data = struct.pack("<h", 32767)
+        samples = pcm16_to_float32(data)
+        assert abs(samples[0] - 32767 / 32768.0) < 1e-5
+
+    def test_float32_to_pcm16_roundtrip(self):
+        original = [0.0, 0.5, -0.5, 0.99, -0.99]
+        pcm = float32_to_pcm16(original)
+        recovered = pcm16_to_float32(pcm)
+        for orig, rec in zip(original, recovered):
+            assert abs(orig - rec) < 1e-3
+
+    def test_float32_to_pcm16_clips(self):
+        # Values > 1.0 should be clipped
+        pcm = float32_to_pcm16([2.0, -2.0])
+        samples = pcm16_to_float32(pcm)
+        assert samples[0] <= 1.0
+        assert samples[1] >= -1.0
+
+
+# ── RMS dB ────────────────────────────────────────────────────────────────────
+
+class TestRmsDb:
+    def test_silence_returns_minus96(self):
+        assert rms_db([]) == -96.0
+        assert rms_db([0.0] * 100) < -60.0
+
+    def test_full_scale_is_near_zero(self):
+        samples = [1.0] * 100
+        db = rms_db(samples)
+        assert abs(db) < 0.1   # 0 dBFS
+
+    def test_half_scale_is_minus6(self):
+        samples = [0.5] * 100
+        db = rms_db(samples)
+        assert abs(db - (-6.02)) < 0.1
+
+    def test_low_level_signal(self):
+        samples = [0.001] * 100
+        db = rms_db(samples)
+        assert db < -40.0
+
+
+# ── Chunk audio ───────────────────────────────────────────────────────────────
+
+class TestChunkAudio:
+    def test_even_chunks(self):
+        samples = list(range(480 * 4))
+        chunks = list(chunk_audio(samples, 480))
+        assert len(chunks) == 4
+        assert all(len(c) == 480 for c in chunks)
+
+    def test_remainder_dropped(self):
+        samples = list(range(500))
+        chunks = list(chunk_audio(samples, 480))
+        assert len(chunks) == 1
+
+    def test_empty_input(self):
+        assert list(chunk_audio([], 480)) == []
+
+
+# ── Energy VAD ────────────────────────────────────────────────────────────────
+
+class TestEnergyVad:
+    def test_silence_not_active(self):
+        vad = EnergyVad(threshold_db=-35.0, onset_frames=2, offset_frames=3)
+        silence = [0.0001] * CHUNK_SAMPLES
+        for _ in range(5):
+            active = vad.process(silence)
+        assert not active
+
+    def test_loud_signal_activates(self):
+        vad = EnergyVad(threshold_db=-35.0, onset_frames=2, offset_frames=8)
+        loud = [0.5] * CHUNK_SAMPLES
+        results = [vad.process(loud) for _ in range(3)]
+        assert results[-1] is True
+
+    def test_onset_requires_n_frames(self):
+        vad = EnergyVad(threshold_db=-35.0, onset_frames=3, offset_frames=8)
+        loud = [0.5] * CHUNK_SAMPLES
+        assert not vad.process(loud)   # frame 1
+        assert not vad.process(loud)   # frame 2
+        assert vad.process(loud)       # frame 3 → activates
+
+    def test_offset_deactivates(self):
+        vad = EnergyVad(threshold_db=-35.0, onset_frames=1, offset_frames=2)
+        loud = [0.5] * CHUNK_SAMPLES
+        silence = [0.0001] * CHUNK_SAMPLES
+        vad.process(loud)
+        assert vad.is_active
+        vad.process(silence)
+        assert vad.is_active   # offset_frames=2, need 2
+        vad.process(silence)
+        assert not vad.is_active
+
+    def test_reset(self):
+        vad = EnergyVad(threshold_db=-35.0, onset_frames=1, offset_frames=8)
+        loud = [0.5] * CHUNK_SAMPLES
+        vad.process(loud)
+        assert vad.is_active
+        vad.reset()
+        assert not vad.is_active
+
+
+# ── Utterance segmenter ───────────────────────────────────────────────────────
+
+class TestUtteranceSegmenter:
+    def _make_segmenter(self):
+        vad = EnergyVad(threshold_db=-35.0, onset_frames=1, offset_frames=2)
+        return UtteranceSegmenter(vad, pre_roll_frames=2, max_duration_s=5.0)
+
+    def test_silence_yields_nothing(self):
+        seg = self._make_segmenter()
+        silence = [0.0001] * CHUNK_SAMPLES
+        for _ in range(10):
+            results = seg.push(silence)
+            assert results == []
+
+    def test_speech_then_silence_yields_utterance(self):
+        seg = self._make_segmenter()
+        loud = [0.5] * CHUNK_SAMPLES
+        silence = [0.0001] * CHUNK_SAMPLES
+
+        # 3 speech frames
+        for _ in range(3):
+            assert seg.push(loud) == []
+
+        # 2 silence frames to trigger offset
+        all_results = []
+        for _ in range(3):
+            all_results.extend(seg.push(silence))
+
+        assert len(all_results) == 1
+        utt_samples, duration = all_results[0]
+        assert duration > 0
+        assert len(utt_samples) > 0
+
+    def test_preroll_included(self):
+        vad = EnergyVad(threshold_db=-35.0, onset_frames=1, offset_frames=2)
+        seg = UtteranceSegmenter(vad, pre_roll_frames=3, max_duration_s=5.0)
+        silence = [0.0001] * CHUNK_SAMPLES
+        loud = [0.5] * CHUNK_SAMPLES
+
+        # Push 3 silence frames (become pre-roll buffer)
+        for _ in range(3):
+            seg.push(silence)
+
+        # Push 1 speech frame (triggers onset, pre-roll included)
+        seg.push(loud)
+
+        # End utterance
+        all_results = []
+        for _ in range(3):
+            all_results.extend(seg.push(silence))
+
+        assert len(all_results) == 1
+        # Utterance should include pre-roll + speech + trailing silence
+        utt_samples, _ = all_results[0]
+        assert len(utt_samples) >= 4 * CHUNK_SAMPLES
+
+
+# ── Speaker identification ────────────────────────────────────────────────────
+
+class TestSpeakerIdentification:
+    def test_cosine_identical(self):
+        v = [1.0, 0.0, 0.0]
+        assert abs(cosine_similarity(v, v) - 1.0) < 1e-6
+
+    def test_cosine_orthogonal(self):
+        a = [1.0, 0.0]
+        b = [0.0, 1.0]
+        assert abs(cosine_similarity(a, b)) < 1e-6
+
+    def test_cosine_opposite(self):
+        a = [1.0, 0.0]
+        b = [-1.0, 0.0]
+        assert abs(cosine_similarity(a, b) + 1.0) < 1e-6
+
+    def test_cosine_zero_vector(self):
+        assert cosine_similarity([0.0, 0.0], [1.0, 0.0]) == 0.0
+
+    def test_cosine_length_mismatch(self):
+        assert cosine_similarity([1.0], [1.0, 0.0]) == 0.0
+
+    def test_identify_speaker_match(self):
+        known = {
+            "alice": [1.0, 0.0, 0.0],
+            "bob":   [0.0, 1.0, 0.0],
+        }
+        result = identify_speaker([1.0, 0.0, 0.0], known, threshold=0.5)
+        assert result == "alice"
+
+    def test_identify_speaker_unknown(self):
+        known = {"alice": [1.0, 0.0, 0.0]}
+        result = identify_speaker([0.0, 1.0, 0.0], known, threshold=0.5)
+        assert result == "unknown"
+
+    def test_identify_speaker_empty_db(self):
+        result = identify_speaker([1.0, 0.0], {}, threshold=0.5)
+        assert result == "unknown"
+
+    def test_identify_best_match(self):
+        known = {
+            "alice": [1.0, 0.0, 0.0],
+            "bob":   [0.8, 0.6, 0.0],  # closer to [1,0,0] than bob is
+        }
+        # Query close to alice
+        result = identify_speaker([0.99, 0.01, 0.0], known, threshold=0.5)
+        assert result == "alice"

jetson/ros2_ws/src/saltybot_social_msgs/CMakeLists.txt

@@ -28,6 +28,11 @@ rosidl_generate_interfaces(${PROJECT_NAME}
   "srv/QueryMood.srv"
   # Issue #92 — multi-modal tracking fusion
   "msg/FusedTarget.msg"
+  # Issue #81 — speech pipeline
+  "msg/SpeechTranscript.msg"
+  "msg/VadState.msg"
+  # Issue #83 — conversation engine
+  "msg/ConversationResponse.msg"
   DEPENDENCIES std_msgs geometry_msgs builtin_interfaces
 )
 

jetson/ros2_ws/src/saltybot_social_msgs/msg/ConversationResponse.msg

@@ -0,0 +1,9 @@
+# ConversationResponse.msg — LLM response, supports streaming token output.
+# Published by conversation_node on /social/conversation/response
+
+std_msgs/Header header
+
+string  text            # Full or partial response text
+string  speaker_id      # Who the response is addressed to
+bool    is_partial       # true = streaming token chunk, false = final response
+int32   turn_id          # Conversation turn counter (for deduplication)

jetson/ros2_ws/src/saltybot_social_msgs/msg/SpeechTranscript.msg

@@ -0,0 +1,10 @@
+# SpeechTranscript.msg — Result of STT with speaker identification.
+# Published by speech_pipeline_node on /social/speech/transcript
+
+std_msgs/Header header
+
+string  text            # Transcribed text (UTF-8)
+string  speaker_id      # e.g. "person_42" or "unknown"
+float32 confidence      # ASR confidence 0..1
+float32 audio_duration  # Duration of the utterance in seconds
+bool    is_partial      # true = intermediate streaming result, false = final

jetson/ros2_ws/src/saltybot_social_msgs/msg/VadState.msg

@@ -0,0 +1,8 @@
+# VadState.msg — Voice Activity Detection state.
+# Published by speech_pipeline_node on /social/speech/vad_state
+
+std_msgs/Header header
+
+bool    speech_active   # true = speech detected, false = silence
+float32 energy_db       # RMS energy in dBFS
+string  state           # "silence" | "speech" | "wake_word"