feat(social): speech pipeline + LLM conversation + TTS + orchestrator (#81 #83 #85 #89)

Issue #81 — Speech pipeline:
- speech_pipeline_node.py: OpenWakeWord "hey_salty" → Silero VAD → faster-whisper
  STT (Orin GPU, <500ms wake-to-transcript) → ECAPA-TDNN speaker diarization
- speech_utils.py: pcm16↔float32, EnergyVad, UtteranceSegmenter (pre-roll, max-
  duration), cosine speaker identification — all pure Python, no ROS2/GPU needed
- Publishes /social/speech/transcript (SpeechTranscript) + /social/speech/vad_state

Issue #83 — Conversation engine:
- conversation_node.py: llama-cpp-python GGUF (Phi-3-mini Q4_K_M, 20 GPU layers),
  streaming token output, per-person sliding-window context (4K tokens), summary
  compression, SOUL.md system prompt, group mode
- llm_context.py: PersonContext, ContextStore (JSON persistence), build_llama_prompt
  (ChatML format), context compression via LLM summarization
- Publishes /social/conversation/response (ConversationResponse, partial + final)

Issue #85 — Streaming TTS:
- tts_node.py: Piper ONNX streaming synthesis, sentence-by-sentence first-chunk
  streaming (<200ms to first audio), sounddevice USB speaker playback, volume control
- tts_utils.py: split_sentences, pcm16_to_wav_bytes, chunk_pcm, apply_volume, strip_ssml

Issue #89 — Pipeline orchestrator:
- orchestrator_node.py: IDLE→LISTENING→THINKING→SPEAKING state machine, GPU memory
  watchdog (throttle at <2GB free), rolling latency stats (p50/p95 per stage),
  VAD watchdog (alert if speech pipeline hangs), /social/orchestrator/state JSON pub
- social_bot.launch.py: brings up all 4 nodes with TimerAction delays

New messages: SpeechTranscript.msg, VadState.msg, ConversationResponse.msg
Config YAMLs: speech_params, conversation_params, tts_params, orchestrator_params
Tests: 58 tests (28 speech_utils + 30 llm_context/tts_utils), all passing

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

jetson/Dockerfile.social

@@ -1,119 +0,0 @@
-# Social-bot container — JetPack 6 + TensorRT + audio AI stack
-# Extends the base ROS2 Humble container with social-bot dependencies.
-#
-# Deps: faster-whisper, llama-cpp-python (CUDA), piper-tts, insightface,
-#       pyannote.audio, OpenWakeWord, Silero VAD, pyaudio, sounddevice
-#
-# Build: docker build -f Dockerfile.social -t saltybot/social:latest .
-# Run:   docker compose -f docker-compose.yml up -d saltybot-social
-
-FROM nvcr.io/nvidia/l4t-jetpack:r36.2.0
-
-LABEL maintainer="sl-jetson"
-LABEL description="Social-bot AI stack — speech, LLM, TTS, face recognition on Orin Nano Super"
-LABEL jetpack="6.0"
-LABEL l4t="r36.2.0"
-
-ENV DEBIAN_FRONTEND=noninteractive
-ENV ROS_DISTRO=humble
-ENV ROS_ROOT=/opt/ros/${ROS_DISTRO}
-ENV PYTHONDONTWRITEBYTECODE=1
-ENV PYTHONUNBUFFERED=1
-
-# ── Locale ────────────────────────────────────────────────────────────────────
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    locales tzdata \
-    && locale-gen en_US.UTF-8 \
-    && rm -rf /var/lib/apt/lists/*
-ENV LANG=en_US.UTF-8
-
-# ── System deps ───────────────────────────────────────────────────────────────
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential cmake git wget curl ca-certificates \
-    python3-dev python3-pip python3-setuptools python3-wheel \
-    # Audio hardware
-    alsa-utils libasound2-dev pulseaudio pulseaudio-utils \
-    portaudio19-dev libsndfile1-dev libsoundio-dev \
-    # USB audio / ReSpeaker support
-    usbutils libusb-1.0-0-dev \
-    # CUDA / TensorRT tools
-    cuda-toolkit-12-2 libcudnn8 libcudnn8-dev \
-    # Misc
-    htop tmux nano ffmpeg \
-    && rm -rf /var/lib/apt/lists/*
-
-# ── ROS2 Humble ───────────────────────────────────────────────────────────────
-RUN curl -sSL https://raw.githubusercontent.com/ros/rosdistro/master/ros.asc \
-    | gpg --dearmor -o /usr/share/keyrings/ros-archive-keyring.gpg && \
-    echo "deb [arch=arm64 signed-by=/usr/share/keyrings/ros-archive-keyring.gpg] \
-    http://packages.ros.org/ros2/ubuntu jammy main" \
-    > /etc/apt/sources.list.d/ros2.list && \
-    apt-get update && apt-get install -y --no-install-recommends \
-    ros-humble-ros-base \
-    python3-colcon-common-extensions \
-    python3-rosdep \
-    && rm -rf /var/lib/apt/lists/*
-
-# ── Python AI deps (core) ─────────────────────────────────────────────────────
-# faster-whisper: CTranslate2 backend, Orin GPU accelerated
-# llama-cpp-python: GGUF quantized LLM, CUDA offload
-# piper-tts: fast neural TTS, CPU/GPU
-# insightface: SCRFD face detection + ArcFace recognition
-# pyannote.audio: speaker diarization + ECAPA-TDNN embeddings
-RUN pip3 install --no-cache-dir \
-    "faster-whisper>=1.0.0" \
-    "ctranslate2>=4.0.0" \
-    "openai-whisper>=20231117" \
-    "piper-tts>=1.2.0" \
-    "insightface>=0.7.3" \
-    "onnxruntime-gpu>=1.17.0" \
-    "pyannote.audio>=3.1.0" \
-    "speechbrain>=1.0.0"
-
-# ── llama-cpp-python with CUDA ─────────────────────────────────────────────────
-# Build with CUDA support for Orin GPU offload
-RUN CMAKE_ARGS="-DLLAMA_CUDA=on -DCUDA_ARCHITECTURES=87" \
-    pip3 install --no-cache-dir llama-cpp-python==0.2.85 --no-binary llama-cpp-python
-
-# ── Wake word / VAD ───────────────────────────────────────────────────────────
-RUN pip3 install --no-cache-dir \
-    "openwakeword>=0.6.0" \
-    "silero-vad>=5.1.0" \
-    "webrtcvad-wheels>=2.0.14"
-
-# ── Audio I/O ─────────────────────────────────────────────────────────────────
-RUN pip3 install --no-cache-dir \
-    "pyaudio>=0.2.14" \
-    "sounddevice>=0.4.6" \
-    "soundfile>=0.12.1" \
-    "numpy>=1.24.0"
-
-# ── TensorRT Python bindings ──────────────────────────────────────────────────
-# Already available via JetPack; install pycuda for custom kernels
-RUN pip3 install --no-cache-dir "pycuda>=2022.2.2"
-
-# ── MQTT for SOUL/agent communication ─────────────────────────────────────────
-RUN pip3 install --no-cache-dir "paho-mqtt>=2.0.0"
-
-# ── ROS2 Python deps ──────────────────────────────────────────────────────────
-RUN pip3 install --no-cache-dir \
-    "rclpy" \
-    "transforms3d>=0.4.1"
-
-# ── Model directory ───────────────────────────────────────────────────────────
-RUN mkdir -p /models/onnx /models/engines /models/gguf /models/piper
-ENV MODEL_DIR=/models
-ENV PIPER_VOICE_DIR=/models/piper
-
-# ── ALSA config for USB mic + speaker ─────────────────────────────────────────
-COPY config/asound.conf /etc/asound.conf
-
-# ── Workspace ─────────────────────────────────────────────────────────────────
-RUN mkdir -p /ros2_ws/src
-WORKDIR /ros2_ws
-
-COPY scripts/entrypoint.sh /entrypoint.sh
-RUN chmod +x /entrypoint.sh
-
-ENTRYPOINT ["/entrypoint.sh"]
-CMD ["bash"]

jetson/config/asound.conf

@@ -1,38 +0,0 @@
-# ALSA config for social-bot audio
-# Assumes USB mic (ReSpeaker or similar) at card 1, USB speaker at card 2
-# Adjust card numbers with: aplay -l / arecord -l
-
-# Default capture device — USB mic array (ReSpeaker 2-Mic or 4-Mic)
-pcm.!default {
-    type asym
-    playback.pcm "usb_speaker"
-    capture.pcm  "usb_mic"
-}
-
-# USB microphone input
-pcm.usb_mic {
-    type plug
-    slave {
-        pcm "hw:1,0"
-        rate 16000
-        channels 1
-        format S16_LE
-    }
-}
-
-# USB speaker output
-pcm.usb_speaker {
-    type plug
-    slave {
-        pcm "hw:2,0"
-        rate 22050
-        channels 1
-        format S16_LE
-    }
-}
-
-# Mixer controls
-ctl.!default {
-    type hw
-    card 1
-}

jetson/docker-compose.yml

@@ -302,61 +302,6 @@ services:
           fence_active:=true
       "
 
-  # ── Social-bot AI stack (speech + LLM + TTS + face recognition) ─────────────
-  # Issue #88: Orin dev environment for social-bot
-  # Start: docker compose up -d saltybot-social
-  # Logs:  docker compose logs -f saltybot-social
-  saltybot-social:
-    image: saltybot/social:latest
-    build:
-      context: .
-      dockerfile: Dockerfile.social
-    container_name: saltybot-social
-    restart: unless-stopped
-    runtime: nvidia
-    network_mode: host
-    depends_on:
-      - stm32-bridge
-    environment:
-      - NVIDIA_VISIBLE_DEVICES=all
-      - NVIDIA_DRIVER_CAPABILITIES=all,audio
-      - ROS_DOMAIN_ID=42
-      - RMW_IMPLEMENTATION=rmw_cyclonedds_cpp
-      - MODEL_DIR=/models
-      - PIPER_VOICE_DIR=/models/piper
-      - LLAMA_N_GPU_LAYERS=20
-      # Audio device routing (override if mic/speaker cards differ)
-      - ALSA_MIC_CARD=1
-      - ALSA_SPEAKER_CARD=2
-    volumes:
-      - ./ros2_ws/src:/ros2_ws/src:rw
-      - ./config:/config:ro
-      # Persistent model storage on NVMe (survives container restarts)
-      - /mnt/nvme/saltybot/models:/models:rw
-      # Enrollment database (face embeddings, voice prints, conversation history)
-      - /mnt/nvme/saltybot/social_db:/social_db:rw
-      # SOUL.md personality file
-      - /mnt/nvme/saltybot/soul:/soul:ro
-    devices:
-      # USB mic array (ReSpeaker 2-Mic / 4-Mic or compatible)
-      - /dev/snd:/dev/snd
-      # USB webcam / face camera (optional, RealSense preferred)
-      - /dev/bus/usb:/dev/bus/usb
-    group_add:
-      - audio
-    cap_add:
-      - SYS_NICE  # allow real-time thread priority for audio
-    ulimits:
-      rtprio: 95
-      memlock: -1
-    command: >
-      bash -c "
-        source /opt/ros/humble/setup.bash &&
-        source /ros2_ws/install/local_setup.bash 2>/dev/null || true &&
-        ros2 launch saltybot_social social_bot.launch.py
-      "
-
-
 volumes:
   saltybot-maps:
     driver: local

jetson/models/README.md

@@ -1,79 +0,0 @@
-# Social-bot Model Directory
-
-## Layout
-
-```
-/models/
-├── onnx/                          # Source ONNX models (version-pinned)
-│   ├── scrfd_10g_bnkps.onnx      # Face detection — InsightFace SCRFD-10GF
-│   ├── arcface_r100.onnx         # Face recognition — ArcFace R100 (buffalo_l)
-│   └── ecapa_tdnn.onnx           # Speaker embedding — ECAPA-TDNN (SpeechBrain export)
-│
-├── engines/                       # TensorRT FP16 compiled engines
-│   ├── scrfd_10g_fp16.engine     # SCRFD → TRT FP16 (640×640)
-│   ├── arcface_r100_fp16.engine  # ArcFace → TRT FP16 (112×112)
-│   └── ecapa_tdnn_fp16.engine    # ECAPA-TDNN → TRT FP16 (variable len)
-│
-├── whisper-small-ct2/             # faster-whisper CTranslate2 format (auto-downloaded)
-│   ├── model.bin
-│   └── tokenizer.json
-│
-├── piper/                         # Piper TTS voice models
-│   ├── en_US-lessac-medium.onnx
-│   └── en_US-lessac-medium.onnx.json
-│
-├── gguf/                          # Quantized LLM (llama-cpp-python)
-│   └── phi-3-mini-4k-instruct-q4_k_m.gguf  # ~2.2GB — Phi-3-mini Q4_K_M
-│
-└── speechbrain_ecapa/             # SpeechBrain pretrained checkpoint cache
-```
-
-## Model Versions
-
-| Model | Version | Source | Size |
-|---|---|---|---|
-| SCRFD-10GF | InsightFace 0.7 | GitHub releases | 17MB |
-| ArcFace R100 (w600k_r50) | InsightFace buffalo_l | Auto via insightface | 166MB |
-| ECAPA-TDNN | SpeechBrain spkrec-ecapa-voxceleb | HuggingFace | 87MB |
-| Whisper small | faster-whisper 1.0+ | CTranslate2 hub | 488MB |
-| Piper en_US-lessac-medium | Rhasspy piper-voices | HuggingFace | 63MB |
-| Phi-3-mini-4k Q4_K_M | microsoft/Phi-3-mini-4k-instruct | GGUF / HuggingFace | 2.2GB |
-
-## Setup
-
-```bash
-# From within the social container:
-/scripts/convert_models.sh all          # download + convert all models
-/scripts/convert_models.sh benchmark    # run latency benchmark suite
-/scripts/convert_models.sh health       # check GPU memory
-```
-
-## Performance Targets (Orin Nano Super, JetPack 6, FP16)
-
-| Model | Input | Target | Typical |
-|---|---|---|---|
-| SCRFD-10GF | 640×640 | <15ms | ~8ms |
-| ArcFace R100 | 4×112×112 | <5ms | ~3ms |
-| ECAPA-TDNN | 1s audio | <20ms | ~12ms |
-| Whisper small | 1s audio | <300ms | ~180ms |
-| Piper lessac-medium | 10 words | <200ms | ~60ms |
-| Phi-3-mini Q4_K_M | prompt | <500ms TTFT | ~350ms |
-
-## LLM Download
-
-```bash
-# Download Phi-3-mini GGUF manually (2.2GB):
-wget -O /models/gguf/phi-3-mini-4k-instruct-q4_k_m.gguf \
-  "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-q4.gguf"
-
-# Or use llama-cpp-python's built-in download:
-python3 -c "
-from llama_cpp import Llama
-llm = Llama.from_pretrained(
-    repo_id='microsoft/Phi-3-mini-4k-instruct-gguf',
-    filename='Phi-3-mini-4k-instruct-q4.gguf',
-    cache_dir='/models/gguf',
-    n_gpu_layers=20
-)
-"
-```

jetson/scripts/convert_models.sh

@@ -1,339 +0,0 @@
-#!/usr/bin/env bash
-# convert_models.sh — Convert social-bot ONNX models to TensorRT FP16 engines
-#
-# Models converted:
-#   SCRFD-10GF       — face detection (InsightFace)
-#   ArcFace-R100     — face recognition (InsightFace)
-#   Whisper-small    — STT via faster-whisper (handled by CTranslate2, not TRT)
-#   ECAPA-TDNN       — speaker embedding (SpeechBrain export)
-#   Piper TTS        — neural TTS (ONNX native, no TRT needed)
-#
-# Usage: ./scripts/convert_models.sh [--model all|scrfd|arcface|ecapa]
-# Requirements: TensorRT 8.6+, trtexec in PATH, ONNX models in /models/onnx/
-#
-# Note: Whisper uses faster-whisper's CTranslate2 which handles its own GPU
-#       optimization. Piper uses ONNX Runtime directly (already optimized).
-#
-set -euo pipefail
-
-MODEL_DIR="${MODEL_DIR:-/models}"
-ONNX_DIR="${MODEL_DIR}/onnx"
-ENGINE_DIR="${MODEL_DIR}/engines"
-WORKSPACE_MB=4096
-TARGET="${1:-all}"
-
-mkdir -p "${ENGINE_DIR}"
-
-log() { echo "[$(date +%H:%M:%S)] $*"; }
-err() { echo "[ERROR] $*" >&2; exit 1; }
-
-check_trtexec() {
-    command -v trtexec &>/dev/null || \
-        err "trtexec not found. Install TensorRT or add to PATH."
-}
-
-# ── Download ONNX sources ──────────────────────────────────────────────────────
-download_models() {
-    log "Downloading ONNX model sources..."
-
-    # SCRFD-10GF (InsightFace model zoo)
-    SCRFD_ONNX="${ONNX_DIR}/scrfd_10g_bnkps.onnx"
-    if [[ ! -f "${SCRFD_ONNX}" ]]; then
-        log "Downloading SCRFD-10GF..."
-        wget -q --show-progress -O "${SCRFD_ONNX}" \
-            "https://github.com/deepinsight/insightface/releases/download/v0.7/scrfd_10g_bnkps.onnx" || \
-        python3 -c "
-import insightface, shutil, os
-app = insightface.app.FaceAnalysis(providers=['CUDAExecutionProvider'])
-app.prepare(ctx_id=0)
-src = os.path.expanduser('~/.insightface/models/buffalo_l/det_10g.onnx')
-if os.path.exists(src): shutil.copy(src, '${SCRFD_ONNX}')
-"
-    fi
-
-    # ArcFace-R100 (InsightFace buffalo_l)
-    ARCFACE_ONNX="${ONNX_DIR}/arcface_r100.onnx"
-    if [[ ! -f "${ARCFACE_ONNX}" ]]; then
-        log "Downloading ArcFace-R100 via insightface..."
-        python3 -c "
-import insightface, shutil, os
-app = insightface.app.FaceAnalysis(providers=['CUDAExecutionProvider'])
-app.prepare(ctx_id=0)
-src = os.path.expanduser('~/.insightface/models/buffalo_l/w600k_r50.onnx')
-if os.path.exists(src): shutil.copy(src, '${ARCFACE_ONNX}')
-"
-    fi
-
-    # ECAPA-TDNN speaker embedding (SpeechBrain pretrained)
-    ECAPA_ONNX="${ONNX_DIR}/ecapa_tdnn.onnx"
-    if [[ ! -f "${ECAPA_ONNX}" ]]; then
-        log "Exporting ECAPA-TDNN to ONNX..."
-        python3 - <<'PYEOF'
-import torch
-from speechbrain.pretrained import EncoderClassifier
-import os
-
-model = EncoderClassifier.from_hparams(
-    source="speechbrain/spkrec-ecapa-voxceleb",
-    savedir="/models/speechbrain_ecapa"
-)
-model.eval()
-
-dummy = torch.randn(1, 16000)  # 1s of audio
-with torch.no_grad():
-    torch.onnx.export(
-        model.encode_batch.__self__,
-        dummy,
-        "/models/onnx/ecapa_tdnn.onnx",
-        opset_version=12,
-        input_names=["waveform"],
-        output_names=["embedding"],
-        dynamic_axes={"waveform": {1: "audio_len"}}
-    )
-print("ECAPA-TDNN exported to /models/onnx/ecapa_tdnn.onnx")
-PYEOF
-    fi
-
-    log "ONNX models ready in ${ONNX_DIR}"
-}
-
-# ── SCRFD face detection ───────────────────────────────────────────────────────
-convert_scrfd() {
-    local ONNX="${ONNX_DIR}/scrfd_10g_bnkps.onnx"
-    local ENGINE="${ENGINE_DIR}/scrfd_10g_fp16.engine"
-
-    [[ -f "${ONNX}" ]] || err "SCRFD ONNX not found at ${ONNX}. Run with --download first."
-
-    log "Converting SCRFD-10GF → TRT FP16..."
-    trtexec \
-        --onnx="${ONNX}" \
-        --saveEngine="${ENGINE}" \
-        --fp16 \
-        --workspace=${WORKSPACE_MB} \
-        --minShapes=input.1:1x3x640x640 \
-        --optShapes=input.1:1x3x640x640 \
-        --maxShapes=input.1:4x3x640x640 \
-        --explicitBatch \
-        2>&1 | tail -5
-
-    log "SCRFD engine saved: ${ENGINE}"
-    benchmark_engine "${ENGINE}" "input.1:1x3x640x640"
-}
-
-# ── ArcFace recognition ────────────────────────────────────────────────────────
-convert_arcface() {
-    local ONNX="${ONNX_DIR}/arcface_r100.onnx"
-    local ENGINE="${ENGINE_DIR}/arcface_r100_fp16.engine"
-
-    [[ -f "${ONNX}" ]] || err "ArcFace ONNX not found at ${ONNX}."
-
-    log "Converting ArcFace-R100 → TRT FP16..."
-    trtexec \
-        --onnx="${ONNX}" \
-        --saveEngine="${ENGINE}" \
-        --fp16 \
-        --workspace=${WORKSPACE_MB} \
-        --minShapes=input.1:1x3x112x112 \
-        --optShapes=input.1:4x3x112x112 \
-        --maxShapes=input.1:16x3x112x112 \
-        --explicitBatch \
-        2>&1 | tail -5
-
-    log "ArcFace engine saved: ${ENGINE}"
-    benchmark_engine "${ENGINE}" "input.1:4x3x112x112"
-}
-
-# ── ECAPA-TDNN speaker embedding ───────────────────────────────────────────────
-convert_ecapa() {
-    local ONNX="${ONNX_DIR}/ecapa_tdnn.onnx"
-    local ENGINE="${ENGINE_DIR}/ecapa_tdnn_fp16.engine"
-
-    [[ -f "${ONNX}" ]] || err "ECAPA-TDNN ONNX not found at ${ONNX}."
-
-    log "Converting ECAPA-TDNN → TRT FP16..."
-    trtexec \
-        --onnx="${ONNX}" \
-        --saveEngine="${ENGINE}" \
-        --fp16 \
-        --workspace=${WORKSPACE_MB} \
-        --minShapes=waveform:1x8000 \
-        --optShapes=waveform:1x16000 \
-        --maxShapes=waveform:1x32000 \
-        --explicitBatch \
-        2>&1 | tail -5
-
-    log "ECAPA-TDNN engine saved: ${ENGINE}"
-    benchmark_engine "${ENGINE}" "waveform:1x16000"
-}
-
-# ── Whisper STT (CTranslate2 — no TRT conversion needed) ─────────────────────
-setup_whisper() {
-    log "Setting up faster-whisper (CTranslate2 format)..."
-    local WHISPER_DIR="/models/whisper-small-ct2"
-
-    if [[ ! -d "${WHISPER_DIR}" ]]; then
-        python3 - <<'PYEOF'
-from faster_whisper import WhisperModel
-import os
-
-# Download and cache the small model in CTranslate2 int8 format
-# This is already GPU-optimized by CTranslate2 — no TRT conversion needed
-model = WhisperModel("small", device="cuda", compute_type="float16",
-                     download_root="/models")
-print("faster-whisper 'small' model ready at /models/")
-# Run one warmup inference
-import numpy as np
-audio = np.zeros(16000, dtype=np.float32)
-segments, _ = model.transcribe(audio)
-list(segments)  # consume generator
-print("Whisper warmup complete")
-PYEOF
-    else
-        log "Whisper already downloaded, skipping."
-    fi
-}
-
-# ── Piper TTS (ONNX Runtime — no TRT conversion needed) ──────────────────────
-setup_piper() {
-    log "Setting up Piper TTS voice models..."
-    local PIPER_DIR="${MODEL_DIR}/piper"
-    mkdir -p "${PIPER_DIR}"
-
-    # Download en_US-lessac-medium (natural, good quality, ~60ms on Orin)
-    if [[ ! -f "${PIPER_DIR}/en_US-lessac-medium.onnx" ]]; then
-        python3 - <<'PYEOF'
-import urllib.request, os
-
-base = "https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/medium/"
-piper_dir = "/models/piper"
-
-for fname in ["en_US-lessac-medium.onnx", "en_US-lessac-medium.onnx.json"]:
-    dest = os.path.join(piper_dir, fname)
-    if not os.path.exists(dest):
-        print(f"Downloading {fname}...")
-        urllib.request.urlretrieve(base + fname, dest)
-        print(f"  → {dest}")
-print("Piper voice ready.")
-PYEOF
-    else
-        log "Piper voice already downloaded, skipping."
-    fi
-}
-
-# ── Benchmark helper ───────────────────────────────────────────────────────────
-benchmark_engine() {
-    local ENGINE="$1"
-    local SHAPE="$2"
-    log "Benchmarking ${ENGINE##*/}..."
-    trtexec \
-        --loadEngine="${ENGINE}" \
-        --shapes="${SHAPE}" \
-        --warmUp=3 \
-        --avgRuns=20 \
-        --useSpinWait \
-        2>&1 | grep -E "mean|median|throughput|Latency" | head -5
-}
-
-# ── Full benchmark suite ───────────────────────────────────────────────────────
-run_benchmarks() {
-    log "=== Social-bot model benchmark suite ==="
-    log "Platform: $(cat /proc/device-tree/model 2>/dev/null || echo 'Orin')"
-    log "TensorRT: $(python3 -c 'import tensorrt as trt; print(trt.__version__)' 2>/dev/null)"
-    log "CUDA: $(nvcc --version 2>/dev/null | grep release | awk '{print $6}' | tr -d ,)"
-    echo
-
-    for ENGINE in "${ENGINE_DIR}"/*.engine; do
-        [[ -f "${ENGINE}" ]] || continue
-        NAME="${ENGINE##*/}"
-        case "${NAME}" in
-            scrfd*)  SHAPE="input.1:1x3x640x640" ;;
-            arcface*)SHAPE="input.1:4x3x112x112" ;;
-            ecapa*)  SHAPE="waveform:1x16000" ;;
-            *)       SHAPE="" ;;
-        esac
-        [[ -z "${SHAPE}" ]] && continue
-        echo "── ${NAME} ──"
-        benchmark_engine "${ENGINE}" "${SHAPE}"
-        echo
-    done
-
-    log "Whisper (CTranslate2 GPU):"
-    python3 - <<'PYEOF'
-import time, numpy as np
-from faster_whisper import WhisperModel
-model = WhisperModel("small", device="cuda", compute_type="float16", download_root="/models")
-audio = np.random.randn(16000).astype(np.float32)
-# warmup
-for _ in range(3):
-    list(model.transcribe(audio)[0])
-# benchmark
-times = []
-for _ in range(10):
-    t = time.perf_counter()
-    list(model.transcribe(audio)[0])
-    times.append((time.perf_counter() - t) * 1000)
-print(f"  1s audio: mean={sum(times)/len(times):.1f}ms, min={min(times):.1f}ms")
-PYEOF
-
-    log "Piper TTS (ONNX Runtime):"
-    python3 - <<'PYEOF'
-import time
-from piper import PiperVoice
-voice = PiperVoice.load("/models/piper/en_US-lessac-medium.onnx")
-text = "Hello, I am Salty, your personal robot companion."
-# warmup
-for _ in range(3):
-    list(voice.synthesize_stream_raw(text))
-# benchmark
-times = []
-for _ in range(10):
-    t = time.perf_counter()
-    audio_bytes = b"".join(voice.synthesize_stream_raw(text))
-    times.append((time.perf_counter() - t) * 1000)
-print(f"  '{text[:30]}...': mean={sum(times)/len(times):.1f}ms, min={min(times):.1f}ms")
-PYEOF
-}
-
-# ── tegrastats snapshot ────────────────────────────────────────────────────────
-check_gpu_health() {
-    log "GPU / system health check:"
-    tegrastats --interval 1000 &
-    TPID=$!
-    sleep 2
-    kill ${TPID} 2>/dev/null
-    echo
-    log "GPU memory:"
-    nvidia-smi --query-gpu=memory.used,memory.free,memory.total,utilization.gpu \
-        --format=csv,noheader,nounits 2>/dev/null || \
-    python3 -c "import pycuda.driver as drv; drv.init(); d=drv.Device(0); c=d.make_context(); print('Free:', drv.mem_get_info()[0]//1024//1024, 'MB'); c.pop()"
-}
-
-# ── Main ───────────────────────────────────────────────────────────────────────
-check_trtexec
-
-case "${TARGET}" in
-    all)
-        download_models
-        convert_scrfd
-        convert_arcface
-        convert_ecapa
-        setup_whisper
-        setup_piper
-        check_gpu_health
-        run_benchmarks
-        ;;
-    scrfd)    download_models; convert_scrfd ;;
-    arcface)  download_models; convert_arcface ;;
-    ecapa)    download_models; convert_ecapa ;;
-    whisper)  setup_whisper ;;
-    piper)    setup_piper ;;
-    benchmark)run_benchmarks ;;
-    health)   check_gpu_health ;;
-    download) download_models ;;
-    *)
-        echo "Usage: $0 [all|scrfd|arcface|ecapa|whisper|piper|benchmark|health|download]"
-        exit 1
-        ;;
-esac
-
-log "Done."

jetson/systemd/install_systemd.sh

@@ -1,33 +0,0 @@
-#!/usr/bin/env bash
-# install_systemd.sh — Install saltybot systemd services on Orin
-# Run as root: sudo ./systemd/install_systemd.sh
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_DIR="$(dirname "${SCRIPT_DIR}")"
-SYSTEMD_DIR="/etc/systemd/system"
-DEPLOY_DIR="/opt/saltybot/jetson"
-
-log() { echo "[install_systemd] $*"; }
-
-[[ "$(id -u)" == "0" ]] || { echo "Run as root"; exit 1; }
-
-# Deploy repo to /opt/saltybot/jetson
-log "Deploying to ${DEPLOY_DIR}..."
-mkdir -p "${DEPLOY_DIR}"
-rsync -a --exclude='.git' --exclude='__pycache__' \
-    "${REPO_DIR}/" "${DEPLOY_DIR}/"
-
-# Install service files
-log "Installing systemd units..."
-cp "${SCRIPT_DIR}/saltybot.target"        "${SYSTEMD_DIR}/"
-cp "${SCRIPT_DIR}/saltybot-social.service" "${SYSTEMD_DIR}/"
-
-# Reload and enable
-systemctl daemon-reload
-systemctl enable saltybot.target
-systemctl enable saltybot-social.service
-
-log "Services installed. Start with:"
-log "  systemctl start saltybot-social"
-log "  journalctl -fu saltybot-social"

jetson/systemd/saltybot-social.service

@@ -1,32 +0,0 @@
-[Unit]
-Description=Saltybot Social-Bot Stack (speech + LLM + TTS + face recognition)
-Documentation=https://gitea.vayrette.com/seb/saltylab-firmware
-After=docker.service network-online.target sound.target
-Requires=docker.service
-PartOf=saltybot.target
-
-[Service]
-Type=simple
-Restart=on-failure
-RestartSec=10s
-TimeoutStartSec=120s
-TimeoutStopSec=30s
-
-User=root
-WorkingDirectory=/opt/saltybot/jetson
-
-# Pull latest image before start (optional — comment out for air-gapped deploy)
-ExecStartPre=-/usr/bin/docker compose -f docker-compose.yml pull saltybot-social
-
-# Start only the social service (not the entire stack)
-ExecStart=/usr/bin/docker compose -f docker-compose.yml up --no-recreate saltybot-social
-
-ExecStop=/usr/bin/docker compose -f docker-compose.yml stop saltybot-social
-
-# Logging
-StandardOutput=journal
-StandardError=journal
-SyslogIdentifier=saltybot-social
-
-[Install]
-WantedBy=saltybot.target multi-user.target

jetson/systemd/saltybot.target

@@ -1,8 +0,0 @@
-[Unit]
-Description=Saltybot Full Stack Target
-Documentation=https://gitea.vayrette.com/seb/saltylab-firmware
-After=docker.service network-online.target
-Requires=docker.service
-
-[Install]
-WantedBy=multi-user.target