tts.py

import os
import torch
import torchaudio
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts
from huggingface_hub import snapshot_download, hf_hub_download
from vinorm import TTSnorm

def generate_speech(text, language="vi", speaker_wav=None):
    # Tải mô hình nếu chưa được tải
    checkpoint_dir = "model/"
    repo_id = "capleaf/viXTTS"
    use_deepspeed = False

    os.makedirs(checkpoint_dir, exist_ok=True)

    required_files = ["model.pth", "config.json", "vocab.json", "speakers_xtts.pth"]
    files_in_dir = os.listdir(checkpoint_dir)
    if not all(file in files_in_dir for file in required_files):
        snapshot_download(
            repo_id=repo_id,
            repo_type="model",
            local_dir=checkpoint_dir,
        )
        hf_hub_download(
            repo_id="coqui/XTTS-v2",
            filename="speakers_xtts.pth",
            local_dir=checkpoint_dir,
        )

    # Cấu hình và tải mô hình
    xtts_config = os.path.join(checkpoint_dir, "config.json")
    config = XttsConfig()
    config.load_json(xtts_config)
    MODEL = Xtts.init_from_config(config)
    MODEL.load_checkpoint(config, checkpoint_dir=checkpoint_dir, use_deepspeed=use_deepspeed)

    if torch.cuda.is_available():
        MODEL.cuda()

    # Chuẩn hóa văn bản
    normalized_text = TTSnorm(text)

    # Tạo giọng nói
    with torch.no_grad():
        gpt_cond_latent, speaker_embedding = MODEL.get_conditioning_latents(audio_path=speaker_wav)
        out = MODEL.inference(
            normalized_text,
            language,
            gpt_cond_latent,
            speaker_embedding,
            temperature=0.7,
        )

    # Lưu file âm thanh
    output_file = "output.wav"
    torchaudio.save(output_file, torch.tensor(out["wav"]).unsqueeze(0), 22050)
    return output_file