Eingebaute Whisper-Instanz: kleiner OpenAI-kompatibler Server (lädt das Modell bei Bedarf, gibt es nach Leerlauf frei)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

deploy/whisper/Dockerfile

@@ -0,0 +1,10 @@
+# diora's built-in Whisper (docs/transcription.md). Built on the server next to diora-prod:
+#   docker compose build whisper
+FROM python:3.12-slim
+RUN apt-get update && apt-get install -y --no-install-recommends ffmpeg && rm -rf /var/lib/apt/lists/*
+RUN pip install --no-cache-dir faster-whisper==1.2.1 fastapi==0.143.1 uvicorn==0.54.0 python-multipart==0.0.32
+WORKDIR /app
+COPY server.py .
+ENV WHISPER_CACHE=/models HF_HOME=/models/hf
+EXPOSE 8000
+CMD ["uvicorn", "server:app", "--host", "0.0.0.0", "--port", "8000"]

deploy/whisper/server.py

@@ -0,0 +1,78 @@
+"""diora's built-in Whisper: a small OpenAI-compatible transcription server.
+
+POST /v1/audio/transcriptions (multipart: file, model, language, prompt, response_format)
+returns {text} or, with response_format=verbose_json, {language, duration, text, segments}.
+The `model` field is ignored; WHISPER_MODEL decides. The model is loaded on the first request
+and freed again after WHISPER_IDLE seconds without work, so it holds no memory while unused.
+One transcription at a time; parallel requests wait (the server has few cores to share).
+See docs/transcription.md.
+"""
+import os
+import subprocess
+import threading
+import time
+
+import numpy as np
+from fastapi import FastAPI, File, Form, HTTPException, UploadFile
+from faster_whisper import WhisperModel
+
+MODEL = os.environ.get('WHISPER_MODEL', 'large-v3-turbo')
+THREADS = int(os.environ.get('WHISPER_THREADS', '4'))
+IDLE = int(os.environ.get('WHISPER_IDLE', '600'))
+CACHE = os.environ.get('WHISPER_CACHE', '/models')
+
+app = FastAPI()
+work = threading.Lock()      # one transcription at a time
+state = {'model': None, 'last': 0.0}
+
+
+def model():
+    if state['model'] is None:
+        state['model'] = WhisperModel(MODEL, device='cpu', compute_type='int8', cpu_threads=THREADS,
+                                      download_root=CACHE)
+    return state['model']
+
+
+def unloader():
+    while True:
+        time.sleep(30)
+        if state['model'] is not None and time.time() - state['last'] > IDLE and work.acquire(blocking=False):
+            try:
+                state['model'] = None
+            finally:
+                work.release()
+
+
+threading.Thread(target=unloader, daemon=True).start()
+
+
+def decode(data: bytes) -> np.ndarray:
+    p = subprocess.run(['ffmpeg', '-nostdin', '-loglevel', 'error', '-i', 'pipe:0', '-ac', '1', '-ar', '16000',
+                        '-f', 'f32le', 'pipe:1'], input=data, capture_output=True)
+    if p.returncode != 0:
+        raise HTTPException(400, 'not audio: ' + p.stderr.decode('utf-8', 'replace')[-200:])
+    return np.frombuffer(p.stdout, np.float32)
+
+
+@app.get('/health')
+def health():
+    return {'ok': True, 'model': MODEL, 'loaded': state['model'] is not None}
+
+
+@app.post('/v1/audio/transcriptions')
+def transcriptions(file: UploadFile = File(...), language: str = Form(''), prompt: str = Form(''),
+                   response_format: str = Form('json')):
+    audio = decode(file.file.read())
+    t0 = time.time()
+    with work:
+        segs, info = model().transcribe(audio, language=language or None, initial_prompt=prompt or None,
+                                        beam_size=5, vad_filter=True)
+        segs = [{'id': i, 'start': round(s.start, 2), 'end': round(s.end, 2), 'text': s.text}
+                for i, s in enumerate(segs)]
+        state['last'] = time.time()
+    dur = len(audio) / 16000
+    print(f'{dur:.0f} s audio in {time.time() - t0:.0f} s', flush=True)
+    text = ''.join(s['text'] for s in segs).strip()
+    if response_format == 'verbose_json':
+        return {'task': 'transcribe', 'language': info.language, 'duration': dur, 'text': text, 'segments': segs}
+    return {'text': text}