Add /api/transcribe endpoint with Whisper

app/services/transcriber.py

@@ -0,0 +1,55 @@
+"""Whisper transcription service — CPU-only, async-ready."""
+
+import io
+import tempfile
+import time
+from pathlib import Path
+
+_model = None
+_model_name = None
+
+
+def _load_model(name: str = "base"):
+    """Lazy-load Whisper model (downloads on first use)."""
+    global _model, _model_name
+    import whisper
+
+    if _model is None or _model_name != name:
+        _model = whisper.load_model(name)
+        _model_name = name
+    return _model
+
+
+def transcribe_bytes(audio_bytes: bytes, model_name: str = "base") -> dict:
+    """Transcribe audio from bytes. Returns {"text": "...", "segments": [...], "language": "..."}"""
+    t0 = time.time()
+
+    model = _load_model(model_name)
+
+    # Write to temp file (whisper needs a file path or numpy array)
+    suffix = ".wav"
+    with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp:
+        tmp.write(audio_bytes)
+        tmp_path = tmp.name
+
+    try:
+        result = model.transcribe(tmp_path, fp16=False)  # fp16=False for CPU
+    finally:
+        Path(tmp_path).unlink(missing_ok=True)
+
+    elapsed = round(time.time() - t0, 1)
+
+    return {
+        "text": result["text"].strip(),
+        "segments": [
+            {
+                "start": round(seg["start"], 2),
+                "end": round(seg["end"], 2),
+                "text": seg["text"].strip(),
+            }
+            for seg in result.get("segments", [])
+        ],
+        "language": result.get("language", "unknown"),
+        "duration_seconds": elapsed,
+        "model": model_name,
+    }