# This demo is adopted from https://github.com/coqui-ai/TTS/blob/dev/TTS/demos/xtts_ft_demo/xtts_demo.py # With some modifications to fit the viXTTS model import argparse import hashlib import logging import os import string import subprocess import sys import tempfile import threading from datetime import datetime import gradio as gr import torch import torchaudio from huggingface_hub import hf_hub_download, snapshot_download from underthesea import sent_tokenize from unidecode import unidecode from TTS.tts.configs.xtts_config import XttsConfig from TTS.tts.models.xtts import Xtts XTTS_MODEL = None SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__)) MODEL_DIR = os.path.join(SCRIPT_DIR, "model") OUTPUT_DIR = os.path.join(SCRIPT_DIR, "output") os.makedirs(OUTPUT_DIR, exist_ok=True) REF_AUDIO_CACHE = {} def clear_gpu_cache(): if torch.cuda.is_available(): torch.cuda.empty_cache() def load_model(checkpoint_dir="model/", repo_id="capleaf/viXTTS", use_deepspeed=False): global XTTS_MODEL clear_gpu_cache() 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): print(f"Missing model files! Downloading from {repo_id}...") 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, ) print( f"Model download finished...") xtts_config = os.path.join(checkpoint_dir, "config.json") config = XttsConfig() config.load_json(xtts_config) XTTS_MODEL = Xtts.init_from_config(config) print( "Loading model...") XTTS_MODEL.load_checkpoint( config, checkpoint_dir=checkpoint_dir, use_deepspeed=False ) if torch.cuda.is_available(): XTTS_MODEL.cuda() else: print("use cpu") XTTS_MODEL.cpu() print("Model Loaded!") return XTTS_MODEL def generate_hash(data): hash_object = hashlib.md5() hash_object.update(data) return hash_object.hexdigest() def get_file_name(text, max_char=50): filename = text[:max_char] filename = filename.lower() filename = filename.replace(" ", "_") filename = filename.translate( str.maketrans("", "", string.punctuation.replace("_", "")) ) filename = unidecode(filename) current_datetime = datetime.now().strftime("%m%d%H%M%S") filename = f"{current_datetime}_{filename}" return filename def normalize_vietnamese_text(text): digits = ["không", "một", "hai", "ba", "bốn", "năm", "sáu", "bảy", "tám", "chín"] text = ( text .replace("..", ".") .replace("!.", "!") .replace("?.", "?") .replace(" .", ".") .replace(" ,", ",") .replace('"', "") .replace("'", "") .replace("AI", "Ây Ai") .replace("A.I", "Ây Ai") ) for i in range(10): text = text.replace(i.__str__(), digits[i]+ " ") return text def calculate_keep_len(text, lang): """Simple hack for short sentences""" if lang in ["ja", "zh-cn"]: return -1 word_count = len(text.split()) num_punct = text.count(".") + text.count("!") + text.count("?") + text.count(",") if word_count < 5: return 15000 * word_count + 2000 * num_punct elif word_count < 10: return 13000 * word_count + 2000 * num_punct return -1 def run_tts(lang, tts_text, speaker_audio_file, normalize_text): global XTTS_MODEL, REF_AUDIO_CACHE if XTTS_MODEL is None: return "Model đang được load. Vui lòng đợi !!", None, None if not speaker_audio_file: return "Cần giọng đọc mẫu !!!", None, None print("Computing conditioning latents...") cache_key_ref_audio = speaker_audio_file if cache_key_ref_audio in REF_AUDIO_CACHE: print("Using conditioning latents cache...") gpt_cond_latent, speaker_embedding = REF_AUDIO_CACHE[cache_key_ref_audio] else: gpt_cond_latent, speaker_embedding = XTTS_MODEL.get_conditioning_latents( audio_path=speaker_audio_file, gpt_cond_len=XTTS_MODEL.config.gpt_cond_len, max_ref_length=XTTS_MODEL.config.max_ref_len, sound_norm_refs=XTTS_MODEL.config.sound_norm_refs, ) REF_AUDIO_CACHE[cache_key_ref_audio] = (gpt_cond_latent, speaker_embedding) tts_text = normalize_vietnamese_text(tts_text) # Split text by sentence if lang in ["ja", "zh-cn"]: sentences = tts_text.split("。") else: sentences = sent_tokenize(tts_text) from pprint import pprint pprint(sentences) wav_chunks = [] for sentence in sentences: if sentence.strip() == "": continue wav_chunk = XTTS_MODEL.inference( text=sentence, language=lang, gpt_cond_latent=gpt_cond_latent, speaker_embedding=speaker_embedding, # The following values are carefully chosen for viXTTS temperature=0.3, length_penalty=1.0, repetition_penalty=10.0, top_k=30, top_p=0.85, enable_text_splitting=True, ) keep_len = calculate_keep_len(sentence, lang) wav_chunk["wav"] = wav_chunk["wav"][:keep_len] wav_chunks.append(torch.tensor(wav_chunk["wav"])) out_wav = torch.cat(wav_chunks, dim=0).unsqueeze(0) gr_audio_id = os.path.basename(os.path.dirname(speaker_audio_file)) out_path = os.path.join(OUTPUT_DIR, f"{get_file_name(tts_text)}_{gr_audio_id}.wav") print("Saving output to ", out_path) torchaudio.save(out_path, out_wav, 24000) return "Speech generated !", out_path # Define a logger to redirect class Logger: def __init__(self, filename="log.out"): self.log_file = filename self.terminal = sys.stdout self.log = open(self.log_file, "w") def write(self, message): self.terminal.write(message) self.log.write(message) def flush(self): self.terminal.flush() self.log.flush() def isatty(self): return False # Redirect stdout and stderr to a file sys.stdout = Logger() sys.stderr = sys.stdout logging.basicConfig( level=logging.ERROR, format="%(asctime)s [%(levelname)s] %(message)s", handlers=[logging.StreamHandler(sys.stdout)], ) def read_logs(): sys.stdout.flush() with open(sys.stdout.log_file, "r") as f: return f.read() def MyThread1(): global XTTS_MODEL XTTS_MODEL = load_model() if __name__ == "__main__": REFERENCE_AUDIO = os.path.join(SCRIPT_DIR, "audio.wav") t1 = threading.Thread(target=MyThread1, args=[]) t1.start() with gr.Blocks() as demo: intro = """ # Fake giọng Demo Customize from HuggingFace: [viXTTS](https://huggingface.co./capleaf/viXTTS) """ gr.Markdown(intro) with gr.Row(): with gr.Column() as col2: speaker_reference_audio = gr.Audio( label="Giọng đọc mẫu:", value=REFERENCE_AUDIO, type="filepath", ) tts_language = gr.Dropdown( label="Language", value="vi", choices=[ "vi", "en", "es", "fr", "de", "it", "pt", "pl", "tr", "ru", "nl", "cs", "ar", "zh", "hu", "ko", "ja", ], ) normalize_text = gr.Checkbox( label="Normalize Input Text", value=True, ) tts_text = gr.Textbox( label="Input Text.", value="Chào bạn, đây là giọng đọc được sinh ra từ AI", ) tts_btn = gr.Button(value="Inference", variant="primary") with gr.Column() as col3: progress_gen = gr.Label(label="Progress:") tts_output_audio = gr.Audio(label="Kết quả.") tts_btn.click( fn=run_tts, inputs=[ tts_language, tts_text, speaker_reference_audio, normalize_text, ], outputs=[progress_gen, tts_output_audio], ) demo.launch()