app.py

# app.py
from fastapi import FastAPI
from fastapi.responses import StreamingResponse
from pydantic import BaseModel
from transformers import VitsModel, AutoTokenizer
import torch, numpy as np, io
from scipy.io.wavfile import write
from fastapi.staticfiles import StaticFiles
import os, pathlib
CACHE = "/tmp/hf-cache"
pathlib.Path(CACHE).mkdir(parents=True, exist_ok=True)
os.environ["HF_HOME"] = os.environ["TRANSFORMERS_CACHE"] = os.environ["XDG_CACHE_HOME"] = CACHE




MODEL = VitsModel.from_pretrained("facebook/mms-tts-bor", cache_dir=CACHE).eval()
TOK = AutoTokenizer.from_pretrained("facebook/mms-tts-bor", cache_dir=CACHE)


app = FastAPI()


class Inp(BaseModel):
    text: str

@app.post("/api/tts")
def tts(inp: Inp):
    with torch.inference_mode():
        audio = MODEL(**TOK(inp.text, return_tensors="pt")).waveform.squeeze().cpu().numpy()
    audio = audio / (np.max(np.abs(audio)) + 1e-9)
    buf = io.BytesIO(); write(buf, MODEL.config.sampling_rate, (audio*32767).astype(np.int16)); buf.seek(0)
    return StreamingResponse(buf, media_type="audio/wav")

app.mount("/", StaticFiles(directory="public", html=True), name="ui")