chore(voice): make NPU Whisper the default

Makefile

@@ -137,23 +137,23 @@ api-dedup: ## Remove duplicate LiteLLM model DB entries.
 api-logs: ## Follow LiteLLM logs.
 	$(DC) logs -f --tail="$(LOGS_TAIL)" litellm litellm-db litellm-init
 
-voice-up: ## Start all voice services.
+voice-up: ## Start default voice services: NPU Whisper and Kokoro TTS.
 	$(DC) --profile voice up -d
 
-voice-gpu: ## Start GPU whisper server and Kokoro TTS.
+voice-gpu: ## Start manual GPU whisper fallback and Kokoro TTS.
-	$(DC) --profile voice up -d whisper-server-gpu kokoro-tts
+	$(DC) --profile voice-gpu --profile voice up -d whisper-server-gpu kokoro-tts
 
 voice-cpu: ## Start CPU whisper server and Kokoro TTS.
-	$(DC) --profile voice up -d whisper-server kokoro-tts
+	$(DC) --profile voice-cpu-backup --profile voice up -d whisper-server kokoro-tts
 
 voice-down: ## Stop voice profile services.
-	$(DC) --profile voice down
+	$(DC) --profile voice --profile voice-gpu --profile voice-cpu-backup down
 
 voice-build: ## Build the custom Blackwell CUDA whisper image.
-	$(DC) --profile voice build whisper-server-gpu
+	$(DC) --profile voice-gpu build whisper-server-gpu
 
-voice-logs: ## Follow voice service logs.
+voice-logs: ## Follow default voice service logs.
-	$(DC) logs -f --tail="$(LOGS_TAIL)" whisper-server-gpu whisper-server kokoro-tts
+	$(DC) logs -f --tail="$(LOGS_TAIL)" whisper-server-npu kokoro-tts
 
 search-up: ## Start Brave Search MCP and SearXNG.
 	$(DC) --profile search up -d

docker-compose.yaml

@@ -37,7 +37,7 @@ services:
   whisper-init:
     image: ghcr.io/ggml-org/whisper.cpp@sha256:672650b5e67f9cb86af7ac6e09dea8eac12a024086e1e5c0172fdccf336aba09
     container_name: whisper-init
-    profiles: ["voice"]
+    profiles: ["voice", "voice-cpu-backup"]
     restart: "no"
     volumes:
       - whisper-models:/app/models
@@ -54,17 +54,15 @@ services:
           fi
         done
 
-  # Primary whisper.cpp server: NVIDIA RTX 5070 Ti via CUDA (Blackwell sm_120).
+  # Manual GPU whisper.cpp fallback: NVIDIA RTX 5070 Ti via CUDA (Blackwell sm_120).
-  # Uses ggml-base.bin to keep the service alive while llama-server owns most of
+  # Kept out of the normal `voice` profile because the OpenVINO NPU Whisper
-  # the laptop GPU VRAM. The previous ggml-small.bin profile needed ~465 MiB
+  # service is the default and this container consumes GPU resources.
-  # contiguous CUDA memory and restarted when only ~560 MiB fragmented VRAM was
-  # free. CPU whisper-server below remains the higher-accuracy fallback.
   #
   # The official `ghcr.io/ggml-org/whisper.cpp:main-cuda` ships kernels only
   # for sm_75/80/86/90 and fails to init CUDA on Blackwell. We build a custom
   # image with `CMAKE_CUDA_ARCHITECTURES=120` from the local Dockerfile.
   # Build manually with: docker build -t whisper.cpp:cuda-blackwell ./whisper-cuda-blackwell
-  # Or `docker compose --profile voice build whisper-server-gpu`.
+  # Or `docker compose --profile voice-gpu build whisper-server-gpu`.
   whisper-server-gpu:
     image: whisper.cpp:cuda-blackwell
     build:
@@ -72,7 +70,7 @@ services:
       dockerfile: Dockerfile
     container_name: whisper-server-gpu
     restart: unless-stopped
-    profiles: ["voice"]
+    profiles: ["voice-gpu"]
     ports:
       - "18801:8080"
     volumes:
@@ -115,16 +113,62 @@ services:
       agentmon.role: "voice"
       agentmon.port: "18801"
 
-  # Fallback whisper.cpp server: CPU-only, medium model.
+  # Experimental OpenVINO GenAI Whisper server using the Intel NPU.
-  # Kept around for resilience — runs if the GPU server is down (driver issue,
+  # This is not whisper.cpp; it implements the same OpenAI-style
-  # gemma takes all VRAM, custom image broken, etc.). Uses no GPU resources.
+  # /v1/audio/transcriptions route using OpenVINO WhisperPipeline on NPU.
-  # ~14 s per short clip (medium-on-CPU is 90x slower than small-on-GPU above).
+  # Host requirements: intel-npu-driver-bin installed, /dev/accel/accel0 present,
-  # Start with: docker compose --profile voice up -d whisper-server
+  # and the host NPU Level Zero driver/compiler libraries mounted below.
+  whisper-server-npu:
+    image: whisper-openvino-npu:local
+    build:
+      context: ./whisper-openvino-npu
+      dockerfile: Dockerfile
+    container_name: whisper-server-npu
+    restart: unless-stopped
+    profiles: ["voice"]
+    ports:
+      - "18816:8080"
+    devices:
+      - /dev/accel/accel0:/dev/accel/accel0
+    group_add:
+      - "987" # host render group gid on willlaptop
+    environment:
+      - WHISPER_DEVICE=NPU
+      - WHISPER_MODEL_DIR=/models/whisper-tiny-fp16-ov
+      - LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu
+      - ZE_ENABLE_ALT_DRIVERS=/usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1
+    volumes:
+      - /home/will/.cache/openvino-models/whisper-tiny-fp16-ov:/models/whisper-tiny-fp16-ov:ro
+      - /usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1.32.1:/usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1.32.1:ro
+      - /usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1:/usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1:ro
+      - /usr/lib/x86_64-linux-gnu/libze_intel_npu.so:/usr/lib/x86_64-linux-gnu/libze_intel_npu.so:ro
+      - /usr/lib/x86_64-linux-gnu/libnpu_driver_compiler.so:/usr/lib/x86_64-linux-gnu/libnpu_driver_compiler.so:ro
+    healthcheck:
+      test:
+        [
+          "CMD-SHELL",
+          "curl -f http://localhost:8080/health >/dev/null 2>&1 || exit 1",
+        ]
+      interval: 30s
+      timeout: 5s
+      start_period: 30s
+      retries: 3
+    labels:
+      agentmon.monitor: "true"
+      agentmon.role: "voice"
+      agentmon.port: "18816"
+
+  # Manual fallback whisper.cpp server: CPU-only, medium model.
+  # Kept around for resilience — runs if the NPU/GPU servers are down. Uses no
+  # accelerator resources, but is slow (~14 s per short clip).
+  # Disabled from the normal `voice` profile now that `whisper-server-npu` is
+  # the trial default. Start manually with:
+  #   docker compose --profile voice-cpu-backup up -d whisper-server
   whisper-server:
     image: ghcr.io/ggml-org/whisper.cpp@sha256:672650b5e67f9cb86af7ac6e09dea8eac12a024086e1e5c0172fdccf336aba09
     container_name: whisper-server
     restart: unless-stopped
-    profiles: ["voice"]
+    profiles: ["voice-cpu-backup"]
     ports:
       - "18811:8080"
     volumes:

docs/swarm-infrastructure.html

@@ -83,7 +83,7 @@
         <!-- Local services -->
         <g><rect x="965" y="165" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="165" width="210" height="80" rx="9" fill="rgba(6,78,59,.4)" stroke="#34d399" stroke-width="1.6"/><text x="1070" y="195" text-anchor="middle" class="title">LiteLLM</text><text x="1070" y="216" text-anchor="middle" class="tiny">LLM router + DB</text><text x="1070" y="234" text-anchor="middle" class="port">:18804</text></g>
         <g><rect x="965" y="275" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="275" width="210" height="80" rx="9" fill="rgba(8,51,68,.4)" stroke="#22d3ee" stroke-width="1.6"/><text x="1070" y="305" text-anchor="middle" class="title">Search</text><text x="1070" y="326" text-anchor="middle" class="tiny">SearXNG + Brave MCP</text><text x="1070" y="344" text-anchor="middle" class="port">:18803 / :18802</text></g>
-        <g><rect x="965" y="385" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="385" width="210" height="80" rx="9" fill="rgba(8,51,68,.4)" stroke="#22d3ee" stroke-width="1.6"/><text x="1070" y="415" text-anchor="middle" class="title">Voice</text><text x="1070" y="436" text-anchor="middle" class="tiny">Kokoro + Whisper</text><text x="1070" y="454" text-anchor="middle" class="port">:18805 / :18811</text></g>
+        <g><rect x="965" y="385" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="385" width="210" height="80" rx="9" fill="rgba(8,51,68,.4)" stroke="#22d3ee" stroke-width="1.6"/><text x="1070" y="415" text-anchor="middle" class="title">Voice</text><text x="1070" y="436" text-anchor="middle" class="tiny">Kokoro + Whisper</text><text x="1070" y="454" text-anchor="middle" class="port">:18805 / :18816</text></g>
         <g><rect x="965" y="555" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="555" width="210" height="80" rx="9" fill="rgba(76,29,149,.4)" stroke="#a78bfa" stroke-width="1.6"/><text x="1070" y="585" text-anchor="middle" class="title">Docker services</text><text x="1070" y="606" text-anchor="middle" class="tiny">agentmon.monitor=true</text><text x="1070" y="624" text-anchor="middle" class="port">swarm/service snapshots</text></g>
         <g><rect x="965" y="665" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="665" width="210" height="80" rx="9" fill="rgba(120,53,15,.3)" stroke="#fbbf24" stroke-width="1.6"/><text x="1070" y="695" text-anchor="middle" class="title">OpenClaw VMs</text><text x="1070" y="716" text-anchor="middle" class="tiny">currently dormant</text><text x="1070" y="734" text-anchor="middle" class="port">openclaw.snapshot</text></g>
         <g><rect x="965" y="775" width="210" height="60" rx="9" fill="#0f172a"/><rect x="965" y="775" width="210" height="60" rx="9" fill="rgba(76,29,149,.4)" stroke="#a78bfa" stroke-width="1.6"/><text x="1070" y="802" text-anchor="middle" class="title">Obsidian / RAG</text><text x="1070" y="822" text-anchor="middle" class="port">:27123/:27124 + ChromaDB</text></g>

docs/swarm-infrastructure.md

@@ -34,7 +34,7 @@ local AI/search/voice services
         +--> llama.cpp :18806
         +--> Ollama embeddings :18807
         +--> Kokoro TTS :18805
-        +--> Whisper :18811
+        +--> Whisper NPU :18816
 ```
 
 See also:
@@ -115,7 +115,7 @@ Docker services:
 - `searxng` — `:18803`, local metasearch
 - `brave-search` — `:18802`, Brave Search MCP server
 - `kokoro-tts` — `:18805`, local TTS
-- `whisper-server` — `:18811`, local transcription
+- `whisper-server-npu` — `:18816`, OpenVINO NPU local transcription
 - `n8n-agent` — `:18808`, automation
 
 Host/user services:

scripts/docker-health-server.py

@@ -24,7 +24,7 @@ CONTAINERS = [
     "litellm-db",
     "n8n-agent",
     "searxng",
-    "whisper-server",
+    "whisper-server-npu",
 ]
 
 

scripts/voice-memo-processor.py

@@ -32,7 +32,7 @@ AUDIO_DIR = os.path.join(tempfile.gettempdir(), "voice-memo-audio")
 os.makedirs(AUDIO_DIR, exist_ok=True)
 
 # Service endpoints (from host perspective)
-WHISPER_URL = os.environ.get("WHISPER_URL", "http://127.0.0.1:18811/v1/audio/transcriptions")
+WHISPER_URL = os.environ.get("WHISPER_URL", "http://127.0.0.1:18816/v1/audio/transcriptions")
 LLM_URL = os.environ.get("LLM_URL", "http://127.0.0.1:18806/v1/chat/completions")
 KOKORO_URL = os.environ.get("KOKORO_URL", "http://127.0.0.1:18805/v1/audio/speech")
 

scripts/voice-memo-service.py

@@ -7,7 +7,7 @@ from http.server import HTTPServer, BaseHTTPRequestHandler
 from pathlib import Path
 
 PORT = int(os.environ.get("VOICE_MEMO_PORT", "18813"))
-WHISPER_URL = os.environ.get("WHISPER_BASE_URL", "http://127.0.0.1:18811")
+WHISPER_URL = os.environ.get("WHISPER_BASE_URL", "http://127.0.0.1:18816")
 LLM_URL = os.environ.get("LLAMA_CPP_BASE_URL", "http://127.0.0.1:18806")
 KOKORO_URL = os.environ.get("KOKORO_BASE_URL", "http://127.0.0.1:18805")
 TELEGRAM_BOT_TOKEN = os.environ.get("TELEGRAM_BOT_TOKEN", "")

whisper-openvino-npu/Dockerfile

@@ -0,0 +1,31 @@
+FROM python:3.14-slim
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1 \
+    LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu \
+    ZE_ENABLE_ALT_DRIVERS=/usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1
+
+RUN apt-get update \
+    && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+       ffmpeg libze1 ca-certificates curl \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN python -m pip install --upgrade pip \
+    && python -m pip install \
+       fastapi==0.126.0 \
+       uvicorn[standard]==0.38.0 \
+       python-multipart==0.0.22 \
+       openvino==2026.2.0 \
+       openvino-genai==2026.2.0.0 \
+       soundfile==0.13.1 \
+       numpy==2.4.6
+
+WORKDIR /app
+COPY server.py /app/server.py
+
+EXPOSE 8080
+HEALTHCHECK --interval=30s --timeout=5s --start-period=30s --retries=3 \
+  CMD curl -fsS http://localhost:8080/health >/dev/null || exit 1
+
+CMD ["uvicorn", "server:app", "--host", "0.0.0.0", "--port", "8080"]

whisper-openvino-npu/server.py

@@ -0,0 +1,147 @@
+import os
+import subprocess
+import tempfile
+import threading
+import time
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import openvino as ov
+import openvino_genai as ov_genai
+import soundfile as sf
+from fastapi import FastAPI, File, Form, UploadFile
+from fastapi.responses import JSONResponse, PlainTextResponse
+
+MODEL_DIR = Path(os.environ.get("WHISPER_MODEL_DIR", "/models/whisper-tiny-fp16-ov"))
+DEVICE = os.environ.get("WHISPER_DEVICE", "NPU")
+BUSY_PATH = Path("/sys/class/accel/accel0/device/npu_busy_time_us")
+
+app = FastAPI(title="OpenVINO NPU Whisper server", version="0.1.0")
+_lock = threading.Lock()
+_pipe = None
+_core = None
+
+
+def busy_us() -> Optional[int]:
+    try:
+        return int(BUSY_PATH.read_text().strip())
+    except Exception:
+        return None
+
+
+def get_core():
+    global _core
+    if _core is None:
+        _core = ov.Core()
+    return _core
+
+
+def get_pipe():
+    global _pipe
+    if _pipe is None:
+        _pipe = ov_genai.WhisperPipeline(str(MODEL_DIR), DEVICE)
+    return _pipe
+
+
+def load_audio(upload_path: Path) -> tuple[np.ndarray, int]:
+    """Decode arbitrary uploaded audio to mono 16 kHz float32 using ffmpeg + soundfile."""
+    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as wav:
+        wav_path = Path(wav.name)
+    try:
+        subprocess.run(
+            [
+                "ffmpeg",
+                "-nostdin",
+                "-hide_banner",
+                "-loglevel",
+                "error",
+                "-y",
+                "-i",
+                str(upload_path),
+                "-ac",
+                "1",
+                "-ar",
+                "16000",
+                "-f",
+                "wav",
+                str(wav_path),
+            ],
+            check=True,
+        )
+        audio, sr = sf.read(wav_path, dtype="float32")
+        if audio.ndim > 1:
+            audio = audio.mean(axis=1)
+        return audio, int(sr)
+    finally:
+        try:
+            wav_path.unlink()
+        except FileNotFoundError:
+            pass
+
+
+@app.get("/")
+def root():
+    return PlainTextResponse("OpenVINO NPU Whisper server\n")
+
+
+@app.get("/health")
+def health():
+    try:
+        core = get_core()
+        devices = core.available_devices
+        npu_name = core.get_property("NPU", "FULL_DEVICE_NAME") if "NPU" in devices else None
+        return {
+            "ok": "NPU" in devices,
+            "device": DEVICE,
+            "devices": devices,
+            "npu": npu_name,
+            "model_dir": str(MODEL_DIR),
+            "model_exists": MODEL_DIR.exists(),
+            "npu_busy_time_us": busy_us(),
+        }
+    except Exception as e:
+        return JSONResponse(status_code=500, content={"ok": False, "error": f"{type(e).__name__}: {e}"})
+
+
+@app.post("/v1/audio/transcriptions")
+async def transcriptions(
+    file: UploadFile = File(...),
+    model: Optional[str] = Form(default=None),
+    language: Optional[str] = Form(default=None),
+    response_format: Optional[str] = Form(default="json"),
+):
+    suffix = Path(file.filename or "audio").suffix or ".audio"
+    with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp:
+        upload_path = Path(tmp.name)
+        tmp.write(await file.read())
+
+    before = busy_us()
+    t0 = time.perf_counter()
+    try:
+        audio, sr = load_audio(upload_path)
+        # OpenVINO GenAI WhisperPipeline appears stateful for Whisper generation on
+        # this stack: reusing one pipeline produced unstable language detection on
+        # repeated short clips. Recreate per request for correctness; OpenVINO's
+        # compiled-cache path keeps warm init reasonably fast.
+        with _lock:
+            pipe = ov_genai.WhisperPipeline(str(MODEL_DIR), DEVICE)
+            result = pipe.generate(audio)
+        text = str(result).strip()
+        elapsed = time.perf_counter() - t0
+        after = busy_us()
+        if response_format == "text":
+            return PlainTextResponse(text)
+        return {
+            "text": text,
+            "duration_seconds": round(elapsed, 4),
+            "sample_rate": sr,
+            "device": DEVICE,
+            "model": model or MODEL_DIR.name,
+            "npu_busy_delta_us": None if before is None or after is None else after - before,
+        }
+    finally:
+        try:
+            upload_path.unlink()
+        except FileNotFoundError:
+            pass