rw_modules/xtts_voice.py

import sys
import io, os, stat
import subprocess
import random
from zipfile import ZipFile
import uuid
import time
import torch
import torchaudio
import langid
import base64
import csv
from io import StringIO
import datetime
import re
from scipy.io.wavfile import write
from pydub import AudioSegment
from TTS.api import TTS
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts
from TTS.utils.generic_utils import get_user_data_dir
from huggingface_hub import HfApi
os.environ["COQUI_TOS_AGREED"] = "1"
repo_id = "coqui/xtts"
from TTS.utils.manage import ModelManager

class xtts_model:
    def __init__(self, model_path):
        device = "cuda:1"
        model_name = "tts_models/multilingual/multi-dataset/xtts_v2"
        config = XttsConfig()
        config.load_json(os.path.join(model_path, "config.json"))
        self.model = Xtts.init_from_config(config)
        self.model.load_checkpoint(config, checkpoint_dir=model_path, eval=True, use_deepspeed=False)
        self.model.to(device).eval()

    def infer(self, prompt, voice, load_sr=20000):
        gpt_cond_latent, speaker_embedding = self.model.get_conditioning_latents(audio_path=voice, gpt_cond_len=30, gpt_cond_chunk_len=4, max_ref_length=60, load_sr=load_sr)
        for out in self.model.inference_stream(prompt,"en", gpt_cond_latent=gpt_cond_latent, speaker_embedding=speaker_embedding, repetition_penalty=5.0, temperature=0.75):
            yield out.to("cpu")
        del gpt_cond_latent, speaker_embedding
        torch.cuda.empty_cache()