Update app.py

app.py

@@ -1,4 +1,4 @@
-import os, io, time
+import os, uuid, tempfile, shutil
 from fastapi import FastAPI, UploadFile, File, HTTPException
 from pydantic import BaseModel
 from faster_whisper import WhisperModel
@@ -6,7 +6,7 @@ from faster_whisper import WhisperModel
 MODEL_NAME = os.getenv("FASTER_WHISPER_MODEL", "tiny.en")
 NUM_THREADS = int(os.getenv("NUM_THREADS", "2"))
 
-# Load model at startup (CPU, int8)
+# Load model once (CPU, int8)
 model = WhisperModel(MODEL_NAME, device="cpu", compute_type="int8", num_workers=NUM_THREADS)
 
 app = FastAPI(title="STT (faster-whisper CPU)")
@@ -16,44 +16,44 @@ class TranscribeOut(BaseModel):
     language: str | None = None
     duration: float | None = None
 
-@app.get("/health")
-def health():
-    return {"ok": True, "model": MODEL_NAME}
-
 @app.post("/transcribe", response_model=TranscribeOut)
-async def transcribe(file: UploadFile = File(...)):
-    # Basic validations
-    if not file.filename:
-        raise HTTPException(status_code=400, detail="No filename")
-    if not file.content_type or not file.content_type.startswith("audio/"):
-        # Allow unknown types; client may not set correctly
-        pass
-
-    # Read all bytes in memory (small test files)
-    try:
-        payload = await file.read()
-    finally:
-        await file.close()
+async def transcribe(
+    file: UploadFile = File(...),
+    beam_size: int = 1,
+    vad_filter: bool = True,
+):
+    # Read the upload
+    payload = await file.read()
+    await file.close()
     if not payload:
         raise HTTPException(status_code=400, detail="Empty file")
 
-    # Run inference
-    start = time.time()
-    audio_buf = io.BytesIO(payload)
-
+    # Persist to a temp file so ffmpeg can probe it robustly
+    suffix = os.path.splitext(file.filename or "")[1] or ".wav"
+    tmp_path = os.path.join(tempfile.gettempdir(), f"stt_{uuid.uuid4().hex}{suffix}")
     try:
-        segments, info = model.transcribe(audio_buf, vad_filter=True)
+        with open(tmp_path, "wb") as f:
+            f.write(payload)
+
+        # Transcribe via file path (lets faster-whisper/ffmpeg do decoding)
+        segments, info = model.transcribe(
+            tmp_path,
+            beam_size=beam_size,
+            vad_filter=vad_filter,
+        )
+
+        parts = [seg.text.strip() for seg in segments if seg.text and seg.text.strip()]
+        text = " ".join(parts).strip()
+
+        return TranscribeOut(
+            text=text,
+            language=getattr(info, "language", None),
+            duration=getattr(info, "duration", None),
+        )
     except Exception as e:
-        # Most common: ffmpeg missing (fixed by Dockerfile), or invalid audio
         raise HTTPException(status_code=500, detail=f"Transcription failed: {e}")
-
-    text_chunks = []
-    for seg in segments:
-        text_chunks.append(seg.text.strip())
-    text = " ".join([t for t in text_chunks if t])
-
-    return TranscribeOut(
-        text=text.strip(),
-        language=getattr(info, "language", None),
-        duration=getattr(info, "duration", None),
-    )
+    finally:
+        try:
+            os.remove(tmp_path)
+        except Exception:
+            pass