import os import numpy as np import torch from torch import no_grad, LongTensor import argparse import commons from mel_processing import spectrogram_torch import utils from models import SynthesizerTrn import gradio as gr import librosa import webbrowser from text import text_to_sequence, _clean_text device = "cuda:0" if torch.cuda.is_available() else "cpu" language_marks = { "Japanese": "", "日本語": "[JA]", "简体中文": "[ZH]", "English": "[EN]", "Mix": "", } lang = ['日本語', '简体中文', 'English', 'Mix'] def get_text(text, hps, is_symbol): text_norm = text_to_sequence( text, hps.symbols, [] if is_symbol else hps.data.text_cleaners) if hps.data.add_blank: text_norm = commons.intersperse(text_norm, 0) text_norm = LongTensor(text_norm) return text_norm def create_tts_fn(model, hps, speaker_ids): def tts_fn(text, speaker, language, speed): if language is not None: text = language_marks[language] + text + language_marks[language] speaker_id = speaker_ids[speaker] stn_tst = get_text(text, hps, False) with no_grad(): x_tst = stn_tst.unsqueeze(0).to(device) x_tst_lengths = LongTensor([stn_tst.size(0)]).to(device) sid = LongTensor([speaker_id]).to(device) audio = model.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=.667, noise_scale_w=0.8, length_scale=1.0 / speed)[0][0, 0].data.cpu().float().numpy() del stn_tst, x_tst, x_tst_lengths, sid return "Success", (hps.data.sampling_rate, audio) return tts_fn def create_vc_fn(model, hps, speaker_ids): def vc_fn(original_speaker, target_speaker, record_audio, upload_audio): input_audio = record_audio if record_audio is not None else upload_audio if input_audio is None: return "You need to record or upload an audio", None sampling_rate, audio = input_audio original_speaker_id = speaker_ids[original_speaker] target_speaker_id = speaker_ids[target_speaker] audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32) if len(audio.shape) > 1: audio = librosa.to_mono(audio.transpose(1, 0)) if sampling_rate != hps.data.sampling_rate: audio = librosa.resample( audio, orig_sr=sampling_rate, target_sr=hps.data.sampling_rate) with no_grad(): y = torch.FloatTensor(audio) y = y / max(-y.min(), y.max()) / 0.99 y = y.to(device) y = y.unsqueeze(0) spec = spectrogram_torch(y, hps.data.filter_length, hps.data.sampling_rate, hps.data.hop_length, hps.data.win_length, center=False).to(device) spec_lengths = LongTensor([spec.size(-1)]).to(device) sid_src = LongTensor([original_speaker_id]).to(device) sid_tgt = LongTensor([target_speaker_id]).to(device) audio = model.voice_conversion(spec, spec_lengths, sid_src=sid_src, sid_tgt=sid_tgt)[0][ 0, 0].data.cpu().float().numpy() del y, spec, spec_lengths, sid_src, sid_tgt return "Success", (hps.data.sampling_rate, audio) return vc_fn if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--model_dir", default="./models/G_9700.pth", help="directory to your fine-tuned model") parser.add_argument("--config_dir", default="./configs/modified_finetune_speaker.json", help="directory to your model config file") parser.add_argument("--share", action="store_true", default=False, help="make link public (used in colab)") args = parser.parse_args() hps = utils.get_hparams_from_file(args.config_dir) net_g = SynthesizerTrn( len(hps.symbols), hps.data.filter_length // 2 + 1, hps.train.segment_size // hps.data.hop_length, n_speakers=hps.data.n_speakers, **hps.model).to(device) _ = net_g.eval() _ = utils.load_checkpoint(args.model_dir, net_g, None) speaker_ids = hps.speakers speakers = list(hps.speakers.keys()) tts_fn = create_tts_fn(net_g, hps, speaker_ids) vc_fn = create_vc_fn(net_g, hps, speaker_ids) app = gr.Blocks() with app: gr.Markdown( "#
vits-fast-fineturning-models-pcr\n" "##
Please do not generate content that could infringe upon the rights or cause harm to individuals or organizations.\n" "##
请不要生成会对个人以及组织造成侵害的内容\n\n" "[![image](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1pn1xnFfdLK63gVXDwV4zCXfVeo8c-I-0?usp=sharing)\n\n" "[![Duplicate this Space](https://huggingface.co/datasets/huggingface/badges/raw/main/duplicate-this-space-sm-dark.svg)](https://huggingface.co/spaces/FrankZxShen/vits-fast-finetuning-pcr?duplicate=true)\n\n" "[![Finetune your own model](https://badgen.net/badge/icon/github?icon=github&label=Finetune%20your%20own%20model)](https://github.com/Plachtaa/VITS-fast-fine-tuning)" ) with gr.Tab("Text-to-Speech"): with gr.Row(): with gr.Column(): textbox = gr.TextArea(label="Text", placeholder="Type your sentence here", value="新たなキャラを解放できるようになったようですね。", elem_id=f"tts-input") # select character char_dropdown = gr.Dropdown( choices=speakers, value=speakers[0], label='character') language_dropdown = gr.Dropdown( choices=lang, value=lang[0], label='language') duration_slider = gr.Slider(minimum=0.1, maximum=5, value=1, step=0.1, label='速度 Speed') with gr.Column(): text_output = gr.Textbox(label="Message") audio_output = gr.Audio( label="Output Audio", elem_id="tts-audio") btn = gr.Button("Generate!") btn.click(tts_fn, inputs=[textbox, char_dropdown, language_dropdown, duration_slider, ], outputs=[text_output, audio_output]) with gr.Tab("Voice Conversion"): gr.Markdown(""" 录制或上传声音，并选择要转换的音色。 """) with gr.Column(): record_audio = gr.Audio( label="record your voice", source="microphone") upload_audio = gr.Audio( label="or upload audio here", source="upload") source_speaker = gr.Dropdown( choices=speakers, value=speakers[0], label="source speaker") target_speaker = gr.Dropdown( choices=speakers, value=speakers[0], label="target speaker") with gr.Column(): message_box = gr.Textbox(label="Message") converted_audio = gr.Audio(label='converted audio') btn = gr.Button("Convert!") btn.click(vc_fn, inputs=[source_speaker, target_speaker, record_audio, upload_audio], outputs=[message_box, converted_audio]) webbrowser.open("http://127.0.0.1:7899") app.queue(concurrency_count=1, api_open=True).launch(share=args.share)