#!/usr/bin/env python3
"""GPU STT service for The Collector — faster-whisper small.en on CUDA (4080S).

POST /transcribe  (raw PCM16 mono 16 kHz body) -> {"text": "..."}
GET  /health
"""
import numpy as np
from fastapi import FastAPI, Request
from faster_whisper import WhisperModel
import uvicorn

MODEL = "small.en"
app = FastAPI()
model = None


@app.on_event("startup")
async def load():
    global model
    model = WhisperModel(MODEL, device="cuda", compute_type="float16")
    print(f"STT ready: {MODEL} on cuda")


@app.get("/health")
async def health():
    return {"ok": model is not None, "model": MODEL, "device": "cuda"}


@app.post("/transcribe")
async def transcribe(request: Request):
    body = await request.body()
    audio = np.frombuffer(body, dtype=np.int16).astype(np.float32) / 32768.0
    segs, _ = model.transcribe(
        audio, beam_size=1, language="en", vad_filter=True,
        condition_on_previous_text=False, no_speech_threshold=0.6)
    text = " ".join(s.text.strip() for s in segs).strip()
    return {"text": text}


if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8770)