python-voiceio 0.2.0__py3-none-any.whl

voiceio/worker.py

@@ -0,0 +1,39 @@
+"""Whisper transcription worker that runs in a subprocess.
+
+Communicates with the parent process via stdin (JSON requests) and
+stdout (JSON responses). Loads the model once, then processes requests
+in a loop for low-latency repeated transcriptions.
+
+Usage: python3 -m voiceio.worker '{"model": "base", ...}'
+"""
+from __future__ import annotations
+
+import base64
+import json
+import sys
+
+import numpy as np
+from faster_whisper import WhisperModel
+
+
+def main() -> None:
+    args = json.loads(sys.argv[1])
+    model = WhisperModel(args["model"], device=args["device"], compute_type=args["compute_type"])
+
+    # Warmup: first transcription is always slow
+    segs, _ = model.transcribe(np.zeros(16000, dtype=np.float32), language=args.get("language"), beam_size=1)
+    list(segs)
+    print("READY", flush=True)
+
+    for line in sys.stdin:
+        if line.strip() == "QUIT":
+            break
+        req = json.loads(line)
+        audio = np.frombuffer(base64.b64decode(req["audio_b64"]), dtype=np.float32)
+        segs, _ = model.transcribe(audio, language=args.get("language"), beam_size=1, best_of=1)
+        text = " ".join(s.text.strip() for s in segs).strip()
+        print(json.dumps({"text": text}), flush=True)
+
+
+if __name__ == "__main__":
+    main()