perf(hermes): use conversational Whisper default

dockerfiles/Dockerfile.hermes-jetson-stt

@@ -14,6 +14,9 @@ RUN python3 -m pip install --no-cache-dir --force-reinstall --no-deps \
 ADD --checksum=sha256:aff26ae408abcba5fbf8813c21e62b0941638c5f6eebfb145be0c9839262a19a --chmod=0444 \
     https://openaipublic.azureedge.net/main/whisper/models/aff26ae408abcba5fbf8813c21e62b0941638c5f6eebfb145be0c9839262a19a/large-v3-turbo.pt \
     /opt/models/whisper/large-v3-turbo.pt
+ADD --checksum=sha256:9ecf779972d90ba49c06d968637d720dd632c55bbf19d441fb42bf17a411e794 --chmod=0444 \
+    https://openaipublic.azureedge.net/main/whisper/models/9ecf779972d90ba49c06d968637d720dd632c55bbf19d441fb42bf17a411e794/small.pt \
+    /opt/models/whisper/small.pt
 RUN chmod 0555 /opt/models /opt/models/whisper
 
 COPY dockerfiles/hermes-jetson-stt-server.py /opt/atlas/hermes-jetson-stt-server.py
@@ -25,11 +28,11 @@ WORKDIR /opt/atlas
 
 # Import the Xavier CUDA stack and confirm Whisper resolves the baked artifact.
 # Full GPU warm-up is covered by the Kubernetes startup probe on titan-21.
-RUN python3 -c "import stat; from pathlib import Path; import torch, whisper; p=Path('/opt/models/whisper'); print(whisper.__file__, whisper.__version__, whisper.available_models()); assert 'large-v3-turbo' in whisper.available_models(); assert stat.S_IMODE(p.stat().st_mode)==0o555; assert stat.S_IMODE((p/'large-v3-turbo.pt').stat().st_mode)==0o444; print(torch.__version__)"
+RUN python3 -c "import stat; from pathlib import Path; import torch, whisper; p=Path('/opt/models/whisper'); print(whisper.__file__, whisper.__version__, whisper.available_models()); assert {'large-v3-turbo','small'} <= set(whisper.available_models()); assert stat.S_IMODE(p.stat().st_mode)==0o555; assert all(stat.S_IMODE((p/name).stat().st_mode)==0o444 for name in ('large-v3-turbo.pt','small.pt')); print(torch.__version__)"
 
 ENV HERMES_STT_HOST=0.0.0.0 \
     HERMES_STT_PORT=9000 \
-    HERMES_STT_MODEL=large-v3-turbo \
+    HERMES_STT_MODEL=small \
     HERMES_STT_CACHE=/opt/models/whisper \
     PYTHONDONTWRITEBYTECODE=1 \
     PYTHONUNBUFFERED=1

dockerfiles/hermes-jetson-stt-server.py

@@ -17,7 +17,7 @@ import whisper
 
 HOST = os.getenv("HERMES_STT_HOST", "0.0.0.0")
 PORT = int(os.getenv("HERMES_STT_PORT", "9000"))
-MODEL_NAME = os.getenv("HERMES_STT_MODEL", "large-v3-turbo")
+MODEL_NAME = os.getenv("HERMES_STT_MODEL", "small")
 CACHE_DIR = Path(os.getenv("HERMES_STT_CACHE", "/cache/whisper"))
 MAX_AUDIO_BYTES = 30 * 1024 * 1024
 MODEL_LOCK = threading.Lock()

services/hermes/chat-configmap.yaml

@@ -39,7 +39,7 @@ data:
       enabled: true
       provider: local_command
       local:
-        model: large-v3-turbo
+        model: small
         language: auto
     model_catalog:
       enabled: true

services/hermes/scripts/hermes_stt_client.py

@@ -48,7 +48,7 @@ def main() -> None:
     parser.add_argument("input_path", type=Path)
     parser.add_argument("--output-dir", required=True, type=Path)
     parser.add_argument("--language", default="auto")
-    parser.add_argument("--model", default="large-v3-turbo")
+    parser.add_argument("--model", default="small")
     args = parser.parse_args()
 
     body, boundary = _multipart(args.input_path, args.language, args.model)

services/hermes/voice-deployment.yaml

@@ -19,7 +19,7 @@ spec:
         app: hermes-stt
       annotations:
         ai.bstein.dev/role: private-chat-speech-to-text
-        ai.bstein.dev/model: whisper-turbo
+        ai.bstein.dev/model: whisper-small-multilingual
         ai.bstein.dev/gpu: titan-21 dedicated speech time-slice
     spec:
       runtimeClassName: nvidia
@@ -31,7 +31,7 @@ spec:
         kubernetes.io/hostname: titan-21
       containers:
         - name: stt
-          image: registry.bstein.dev/bstein/hermes-jetson-stt@sha256:d70bc0892bfce7dc474b24c62d2427a22c6022d0d406c349fcb83d82130de768
+          image: registry.bstein.dev/bstein/hermes-jetson-stt@sha256:341fd5c93d202d7368bbd74bd392f207c66e6d3968cc39f2d1684ec87a9d3899
           imagePullPolicy: IfNotPresent
           ports:
             - {name: http, containerPort: 9000, protocol: TCP}
@@ -39,7 +39,7 @@ spec:
             - {name: HOME, value: /tmp}
             - {name: XDG_CACHE_HOME, value: /tmp/cache}
             - {name: HERMES_STT_PORT, value: "9000"}
-            - {name: HERMES_STT_MODEL, value: large-v3-turbo}
+            - {name: HERMES_STT_MODEL, value: small}
             - {name: HERMES_STT_CACHE, value: /opt/models/whisper}
             - {name: NVIDIA_VISIBLE_DEVICES, value: all}
             - {name: NVIDIA_DRIVER_CAPABILITIES, value: compute,utility}

testing/tests/test_hermes_chat_quality.py

@@ -171,7 +171,7 @@ def test_chat_voice_uses_private_jetson_services_and_shared_auto_route():
     assert config["stt"] == {
         "enabled": True,
         "provider": "local_command",
-        "local": {"model": "large-v3-turbo", "language": "auto"},
+        "local": {"model": "small", "language": "auto"},
     }
 
     dockerfile = (ROOT / "dockerfiles" / "Dockerfile.hermes-webui").read_text()
@@ -191,6 +191,7 @@ def test_voice_models_are_baked_and_runtime_has_no_public_egress():
     stt_dockerfile = (ROOT / "dockerfiles" / "Dockerfile.hermes-jetson-stt").read_text()
     tts_dockerfile = (ROOT / "dockerfiles" / "Dockerfile.hermes-jetson-tts").read_text()
     assert "ADD --checksum=sha256:aff26ae4" in stt_dockerfile
+    assert "ADD --checksum=sha256:9ecf7799" in stt_dockerfile
     assert "--chmod=0444" in stt_dockerfile
     assert "chmod 0555 /opt/models /opt/models/whisper" in stt_dockerfile
     assert "HERMES_STT_CACHE=/opt/models/whisper" in stt_dockerfile