VITS-fast-fine-tuning

This repo is a pipeline of VITS finetuning for fast speaker adaptation TTS, and many-to-many voice conversion

4,870

725

4,870

349

View on GitHub

Top Related Projects

StyleTTS2

5,682

StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models

Retrieval-based-Voice-Conversion-WebUI

28,362

Easily train a good VC model with voice data <= 10 mins!

Quick Overview

VITS-fast-fine-tuning is a GitHub repository that provides a streamlined approach to fine-tuning VITS (Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech) models. It offers a user-friendly interface and optimized training process, allowing users to create custom voice models with minimal data and computational resources.

Pros

Fast fine-tuning process, requiring only 5-10 minutes of training data
User-friendly interface with Google Colab integration for easy setup and execution
Supports multiple languages and can generate multi-lingual results
Includes pre-trained models and datasets for quick experimentation

Cons

Limited documentation for advanced customization and troubleshooting
Dependency on specific versions of libraries, which may cause compatibility issues
Potential overfitting on small datasets, affecting the quality of generated speech
Limited control over the fine-tuning process compared to more comprehensive TTS frameworks

Code Examples

Loading a pre-trained model and generating speech:

from vits import utils, commons
from vits.models import SynthesizerTrn
import torch

hps = utils.get_hparams_from_file("configs/ljs_base.json")
net_g = SynthesizerTrn(
    len(hps.symbols),
    hps.data.filter_length // 2 + 1,
    hps.train.segment_size // hps.data.hop_length,
    **hps.model)
_ = net_g.eval()

_ = utils.load_checkpoint("pretrained_ljs.pth", net_g, None)

text = "Hello, this is a test."
stn_tst = utils.get_text(text, hps)
with torch.no_grad():
    x_tst = stn_tst.unsqueeze(0)
    x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
    audio = net_g.infer(x_tst, x_tst_lengths, noise_scale=.667, noise_scale_w=0.8, length_scale=1)[0][0,0].data.cpu().float().numpy()

Fine-tuning a model with custom data:

from vits import train
import json

config_path = "configs/custom_config.json"
with open(config_path, "r") as f:
    hparams = json.load(f)

train.train(hparams, "custom_output_directory")

Inference using a fine-tuned model:

from vits import utils, commons
from vits.models import SynthesizerTrn
import torch

hps = utils.get_hparams_from_file("custom_output_directory/config.json")
net_g = SynthesizerTrn(
    len(hps.symbols),
    hps.data.filter_length // 2 + 1,
    hps.train.segment_size // hps.data.hop_length,
    **hps.model)
_ = net_g.eval()

_ = utils.load_checkpoint("custom_output_directory/G_latest.pth", net_g, None)

text = "This is a test of the fine-tuned model."
stn_tst = utils.get_text(text, hps)
with torch.no_grad():
    x_tst = stn_tst.unsqueeze(0)
    x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
    audio = net_g.infer(x_tst, x_tst_lengths, noise_scale=.667, noise_scale_w=0.8, length_scale=1)[0][0,0].data.cpu().float().numpy()

Getting Started

Clone the repository:

git clone https://github.com/Plachtaa/VITS-fast-fine-tuning.git
cd VITS-fast-fine-tuning

Competitor Comparisons

StyleTTS2

5,682

StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models

Pros of StyleTTS2

Offers more advanced style control and transfer capabilities
Supports multi-speaker TTS with better voice cloning abilities
Provides improved audio quality and naturalness in generated speech

Cons of StyleTTS2

Requires more computational resources for training and inference
Has a steeper learning curve due to its more complex architecture
May have longer processing times compared to VITS-fast-fine-tuning

Code Comparison

StyleTTS2:

output = model.infer(text, speaker_embedding, style_embedding)

VITS-fast-fine-tuning:

stn_tst = get_text(text, hps)
x_tst = stn_tst.cuda().unsqueeze(0)
x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).cuda()
audio = net_g.infer(x_tst, x_tst_lengths, noise_scale=.667, noise_scale_w=0.8, length_scale=1)[0][0,0].data.cpu().float().numpy()

StyleTTS2 offers a more streamlined inference process, while VITS-fast-fine-tuning requires more manual setup and parameter adjustments. StyleTTS2's code is more concise and easier to use, but may hide some of the underlying complexity. VITS-fast-fine-tuning provides more direct control over the inference parameters, which can be beneficial for fine-tuning the output.

Retrieval-based-Voice-Conversion-WebUI

28,362

Easily train a good VC model with voice data <= 10 mins!

Pros of Retrieval-based-Voice-Conversion-WebUI

Offers a user-friendly web interface for voice conversion tasks
Supports real-time voice conversion with low latency
Provides a wider range of pre-trained models and voice datasets

Cons of Retrieval-based-Voice-Conversion-WebUI

May require more computational resources for real-time processing
Less focused on fine-tuning specific voices compared to VITS-fast-fine-tuning
Potentially more complex setup due to additional dependencies

Code Comparison

VITS-fast-fine-tuning:

def train(rank, a, h):
    if h.num_gpus > 1:
        init_process_group(backend=h.dist_config['dist_backend'], init_method=h.dist_config['dist_url'], world_size=h.num_gpus, rank=rank)

Retrieval-based-Voice-Conversion-WebUI:

def vc_single(
    sid,
    input_audio,
    f0_up_key,
    f0_method,
    file_index,
    index_rate,
    filter_radius,
    resample_sr,
    rms_mix_rate,
    protect,
):
    # Voice conversion logic

The code snippets show different approaches: VITS-fast-fine-tuning focuses on distributed training setup, while Retrieval-based-Voice-Conversion-WebUI emphasizes the voice conversion process with various parameters for customization.

Convert designs to code with AI

Introducing Visual Copilot: A new AI model to turn Figma designs to high quality code using your components.

Try Visual Copilot

README

ä¸æææ¡£è¯·ç¹å»è¿é

VITS Fast Fine-tuning

This repo will guide you to add your own character voices, or even your own voice, into existing VITS TTS model to make it able to do the following tasks in less than 1 hour:

Many-to-many voice conversion between any characters you added & preset characters in the model.
English, Japanese & Chinese Text-to-Speech synthesis with the characters you added & preset characters

Welcome to play around with the base models!
Chinese & English & Japaneseï¼ Author: Me

Chinese & Japaneseï¼ Author: SayaSS

Chinese onlyï¼(No running huggingface spaces) Author: Wwwwhy230825

Currently Supported Tasks:

Clone character voice from 10+ short audios
Clone character voice from long audio(s) >= 3 minutes (one audio should contain single speaker only)
Clone character voice from videos(s) >= 3 minutes (one video should contain single speaker only)
Clone character voice from BILIBILI video links (one video should contain single speaker only)

Currently Supported Characters for TTS & VC:

Any character you wish as long as you have their voices! (Note that voice conversion can only be conducted between any two speakers in the model)

Fine-tuning

See LOCAL.md for local training guide.
Alternatively, you can perform fine-tuning on Google Colab

How long does it take?

Install dependencies (3 min)
Choose pretrained model to start. The detailed differences between them are described in Colab Notebook
Upload the voice samples of the characters you wish to addï¼see DATA.MD for detailed uploading options.
Start fine-tuning. Time taken varies from 20 minutes ~ 2 hours, depending on the number of voices you uploaded.

Inference or Usage (Currently support Windows only)

Remember to download your fine-tuned model!
Download the latest release
Put your model & config file into the folder inference, which are named G_latest.pth and finetune_speaker.json, respectively.
The file structure should be as follows:

inference
ââââinference.exe
ââââ...
ââââfinetune_speaker.json
ââââG_latest.pth

run inference.exe, the browser should pop up automatically.
Note: you must install ffmpeg to enable voice conversion feature.

Inference with CLI

In this example, we will show how to run inference with the default pretrained model. We are now in the main repository directory.

Create the necessary folders and download the necessary files.

cd monotonic_align/
mkdir monotonic_align
python setup.py build_ext --inplace
cd ..
mkdir pretrained_models
# download data for fine-tuning
wget https://huggingface.co/datasets/Plachta/sampled_audio4ft/resolve/main/sampled_audio4ft_v2.zip
unzip sampled_audio4ft_v2.zip

For your finetuned model you may need to create additional directories:

mkdir video_data
mkdir raw_audio
mkdir denoised_audio
mkdir custom_character_voice
mkdir segmented_character_voice

Download pretrained models. For example, trilingual model:

wget https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer/resolve/main/pretrained_models/D_trilingual.pth -O ./pretrained_models/D_0.pth
wget https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer/resolve/main/pretrained_models/G_trilingual.pth -O ./pretrained_models/G_0.pth
wget https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer/resolve/main/configs/uma_trilingual.json -O ./configs/finetune_speaker.json

Activate your environment and run the following code: python3 cmd_inference.py -m pretrained_models/G_0.pth -c configs/finetune_speaker.json -t ä½ å¥½ï¼è®ç»ååçï¼å¾é«å´è§å°ä½ ã -s "æ´¾è Paimon (Genshin Impact)" -l "ç®ä½ä¸æ" You can choose another language, customize output folder, change text and character, but all these parameters you can see in the file cmd_inference.py. Below I'll show only how to change the character.
To change the character please open config file (configs/finetune_speaker.json). There you can find dictionary speakers, where you'll be able to see full list of speakers. Just copy the name of the character you need use it instead of "æ´¾è Paimon (Genshin Impact)"
If you have success, you can find output .wav file in the output/vits

Use in MoeGoe

Prepare downloaded model & config file, which are named G_latest.pth and moegoe_config.json, respectively.
Follow MoeGoe page instructions to install, configure path, and use.

Looking for help?

If you have any questions, please feel free to open an issue or join our Discord server.

Top Related Projects

StyleTTS2

5,682

StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models

Retrieval-based-Voice-Conversion-WebUI

28,362

Easily train a good VC model with voice data <= 10 mins!

Convert designs to code with AI

Introducing Visual Copilot: A new AI model to turn Figma designs to high quality code using your components.

Try Visual Copilot