stt-server: FastAPI whisper.cpp Vulkan ASR + pyannote diarization, containerizzato

server.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+"""STT Server persistente (frigate.vpn:8883).
+ASR: whisper.cpp (Vulkan) | Diarization opzionale: pyannote Community-1 (CPU)
+
+Endpoints:
+  GET  /health            → {"status": "ok", "model": ...}
+  POST /transcribe        → multipart: file=<audio> + form: lang, model, diarize
+      diarize=false (default): ASR only → {"text": "...", "segments": [...]}
+      diarize=true:            ASR + diarization → segments con speaker
+
+Uso: uvicorn server:app --host 0.0.0.0 --port 8883
+"""
+import asyncio, os, tempfile, time
+from pathlib import Path
+
+from fastapi import FastAPI, File, Form, UploadFile
+from fastapi.responses import JSONResponse
+
+import transcribe
+
+app = FastAPI(title="STT Server", version="1.0")
+
+# Semaforo: whisper-cli Vulkan e pyannote non sono thread-safe per uso concorrente
+_lock = asyncio.Lock()
+
+# Cache dei modelli già caricati (pyannote pipeline è costosa da caricare)
+_pipeline_cache = {}
+
+
+def _asr_only(audio_path: str, lang: str, model: str) -> list[dict]:
+    """Path veloce: solo ASR (whisper.cpp Vulkan), niente diarization."""
+    with tempfile.TemporaryDirectory() as tmp:
+        segs = transcribe.asr(audio_path, model, lang, tmp)
+    return segs
+
+
+def _full_pipeline(audio_path: str, lang: str, model: str) -> list[dict]:
+    """Path completo: ASR + diarization pyannote."""
+    with tempfile.TemporaryDirectory() as tmp:
+        segs = transcribe.asr(audio_path, model, lang, tmp)
+        turns = transcribe.diarize(audio_path, tmp)
+        segs = transcribe.merge(segs, turns)
+    return segs
+
+
+@app.get("/health")
+async def health():
+    return {"status": "ok", "service": "stt", "model": "large-v3-turbo"}
+
+
+@app.post("/transcribe")
+async def transcribe_endpoint(
+    file: UploadFile = File(...),
+    lang: str = Form("it"),
+    model: str = Form("large-v3-turbo"),
+    diarize: bool = Form(False),
+):
+    t0 = time.time()
+    # Salva l'upload in un file temporaneo
+    suffix = Path(file.filename or "audio.wav").suffix or ".wav"
+    with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp:
+        tmp.write(await file.read())
+        audio_path = tmp.name
+    try:
+        async with _lock:
+            loop = asyncio.get_running_loop()
+            if diarize:
+                segs = await loop.run_in_executor(None, _full_pipeline, audio_path, lang, model)
+            else:
+                segs = await loop.run_in_executor(None, _asr_only, audio_path, lang, model)
+    finally:
+        os.unlink(audio_path)
+
+    text = " ".join(s["text"] for s in segs).strip()
+    elapsed = round(time.time() - t0, 2)
+    return JSONResponse({
+        "text": text,
+        "segments": segs,
+        "diarize": diarize,
+        "lang": lang,
+        "model": model,
+        "elapsed_s": elapsed,
+    })
+
+
+if __name__ == "__main__":
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=int(os.environ.get("PORT", "8883")))