chore(voice): make NPU Whisper the default

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).
-  # Uses ggml-base.bin to keep the service alive while llama-server owns most of
-  # the laptop GPU VRAM. The previous ggml-small.bin profile needed ~465 MiB
-  # contiguous CUDA memory and restarted when only ~560 MiB fragmented VRAM was
-  # free. CPU whisper-server below remains the higher-accuracy fallback.
+  # Manual GPU whisper.cpp fallback: NVIDIA RTX 5070 Ti via CUDA (Blackwell sm_120).
+  # Kept out of the normal `voice` profile because the OpenVINO NPU Whisper
+  # service is the default and this container consumes GPU resources.
   #
   # 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.
-  # Kept around for resilience — runs if the GPU server is down (driver issue,
-  # gemma takes all VRAM, custom image broken, etc.). Uses no GPU resources.
-  # ~14 s per short clip (medium-on-CPU is 90x slower than small-on-GPU above).
-  # Start with: docker compose --profile voice up -d whisper-server
+  # Experimental OpenVINO GenAI Whisper server using the Intel NPU.
+  # This is not whisper.cpp; it implements the same OpenAI-style
+  # /v1/audio/transcriptions route using OpenVINO WhisperPipeline on NPU.
+  # Host requirements: intel-npu-driver-bin installed, /dev/accel/accel0 present,
+  # 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: