app.py

import os
import io
from fastapi import FastAPI
from pydantic import BaseModel
from transformers import AutoProcessor, VitsForConditionalGeneration
import torch
from fastapi.responses import StreamingResponse

# Use /tmp for cache to avoid permission errors
os.environ["HF_HOME"] = "/tmp"

app = FastAPI()

# Load processor and model once on startup
model_name = "Somali-tts/somali_tts_model"
processor = AutoProcessor.from_pretrained(model_name)
model = VitsForConditionalGeneration.from_pretrained(model_name)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

class TextInput(BaseModel):
    inputs: str

@app.post("/synthesize")
async def synthesize_tts(data: TextInput):
    inputs = processor(data.inputs, return_tensors="pt").to(device)
    with torch.no_grad():
        audio = model.generate(**inputs)
    audio = audio.squeeze().cpu().numpy()

    # Convert to WAV bytes in-memory
    import soundfile as sf
    buf = io.BytesIO()
    sf.write(buf, audio, samplerate=22050, format="WAV")
    buf.seek(0)

    return StreamingResponse(buf, media_type="audio/wav")