app.py

import io
import os
import time

from fastapi import FastAPI, File, UploadFile
from faster_whisper import WhisperModel

# The model is configurable via an environment variable so you can trade
# speed vs. accuracy without touching code:
#     heroku config:set WHISPER_MODEL=tiny.en    (fastest)
#     heroku config:set WHISPER_MODEL=base.en    (default, balanced)
#     heroku config:set WHISPER_MODEL=small.en   (most accurate, slowest)
MODEL_NAME = os.environ.get("WHISPER_MODEL", "base.en")

# int8 keeps memory + latency low on a CPU dyno (Heroku has no GPU).
# The model downloads once when the dyno starts, then stays cached for the
# life of that dyno. The very first request after a deploy/restart is slow
# because of that one-time download + load.
model = WhisperModel(MODEL_NAME, device="cpu", compute_type="int8")

app = FastAPI(title="Voice STT server")


@app.get("/")
def health():
    """Simple health check so you can confirm the dyno is up."""
    return {"status": "ok", "model": MODEL_NAME}


@app.post("/transcribe")
async def transcribe(file: UploadFile = File(...)):
    """Accepts an audio file, returns the transcribed text.

    Test from a computer:
        curl -X POST -F "file=@sample.wav" https://YOUR-APP.herokuapp.com/transcribe
    """
    audio_bytes = await file.read()

    started = time.time()
    segments, _info = model.transcribe(
        io.BytesIO(audio_bytes),
        language="en",
        beam_size=1,        # greedy decoding = fastest
        vad_filter=True,    # trims silence so short clips come back quicker
    )
    text = " ".join(segment.text.strip() for segment in segments).strip()
    elapsed = round(time.time() - started, 2)

    return {"text": text, "seconds": elapsed, "model": MODEL_NAME}