tts_engine.py

import torch
import soundfile as sf
import tempfile

from transformers import AutoTokenizer
from transformers import AutoModelForSeq2SeqLM

model = None
tokenizer = None


def load_model():

    global model, tokenizer

    if model is None:

        model_name = "sarvamai/sarvam-tts"

        tokenizer = AutoTokenizer.from_pretrained(model_name)

        model = AutoModelForSeq2SeqLM.from_pretrained(model_name)


def generate_voice(text):

    load_model()

    inputs = tokenizer(text, return_tensors="pt")

    with torch.no_grad():

        audio = model.generate(**inputs)

    audio_file = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")

    sf.write(audio_file.name, audio.numpy(), 22050)

    return audio_file.name