又发现一个AI语音克隆“降维打击”……

之前分享的 CosyVoice 3秒可克隆，觉得已经够吓人了，结果今天这个更要命，随便录了1分钟自己的声音训练后，它直接把声线、语气、情感、呼吸、停顿全部复刻，简直像本人灵魂附体！

阿里达摩院的 CosyVoice 刚21.2k星，而这款已狂飙至57.8k星，成为开源语音克隆领域的绝对王者！

这是啥工具？

叫 GPT-SoVITS，最大特点就是超强少样本 + 高保真声音克隆：

零样本：只需5秒声音，瞬间TTS，开箱即用

少样本微调：1分钟录音训练后，相似度、自然度和情感表现直接拉满，远超CosyVoice

支持跨语言（中文训练后直接说英语、日语、韩语、粤语等，声线完全不变）

自带完整WebUI工具链：人声分离 → 自动切分 → ASR标注 → 一键训练 → 推理，小白全程鼠标点点就能玩转

开源免费（MIT协议），本地运行，零上传，隐私安全

CosyVoice胜在“3秒即用”的极简，但GPT-SoVITS在1分钟训练后的真实感、情感丰富度和长时间稳定性上实现降维打击，尤其适合需要高保真输出的重度用户。

现在AI声音克隆已经卷到这个地步，真的彻底疯了！

：https://github.com/RVC-Boss/GPT-SoVITS…

强烈推荐大家试一把，体验真正的声音克隆天花板！

记得只克隆自己或获得授权的声音，合规使用～

#AI #工具分享 #GPTSoVITS #语音克隆 #开源神器

RVC-Boss/GPT-SoVITS

Source: https://github.com/RVC-Boss/GPT-SoVITS

Features:

1. Zero-shot TTS: Input a 5-second vocal sample and experience instant text-to-speech conversion.

2. Few-shot TTS: Fine-tune the model with just 1 minute of training data for improved voice similarity and realism.

3. Cross-lingual Support: Inference in languages different from the training dataset, currently supporting English, Japanese, Korean, Cantonese and Chinese.

4. WebUI Tools: Integrated tools include voice accompaniment separation, automatic training set segmentation, Chinese ASR, and text labeling, assisting beginners in creating training datasets and GPT/SoVITS models.

Check out our demo video here!

Unseen speakers few-shot fine-tuning demo:

https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb

RTF(inference speed) of GPT-SoVITS v2 ProPlus: 0.028 tested in 4060Ti, 0.014 tested in 4090 (1400words~=4min, inference time is 3.36s), 0.526 in M4 CPU. You can test our huggingface demo (half H200) to experience high-speed inference .

请不要尬黑GPT-SoVITS推理速度慢，谢谢！

CPU-Optimized Inference Version：https://github.com/baicai-1145/GPT-SoVITS-CPUFast

User guide: 简体中文 | English

Installation

For users in China, you can click here to use AutoDL Cloud Docker to experience the full functionality online.

Tested Environments

Windows

If you are a Windows user (tested with win>=10), you can download the integrated package and double-click on go-webui.bat to start GPT-SoVITS-WebUI.

Users in China can download the package here.

Install the program by running the following commands:

conda create -n GPTSoVits python=3.10
conda activate GPTSoVits
pwsh -F install.ps1 --Device <CU126|CU128|CPU> --Source <HF|HF-Mirror|ModelScope> [--DownloadUVR5]

Linux

conda create -n GPTSoVits python=3.10
conda activate GPTSoVits
bash install.sh --device <CU126|CU128|ROCM|CPU> --source <HF|HF-Mirror|ModelScope> [--download-uvr5]

macOS

Note: The models trained with GPUs on Macs result in significantly lower quality compared to those trained on other devices, so we are temporarily using CPUs instead.

Install the program by running the following commands:

conda create -n GPTSoVits python=3.10
conda activate GPTSoVits
bash install.sh --device <MPS|CPU> --source <HF|HF-Mirror|ModelScope> [--download-uvr5]

Install Manually

Install Dependences

conda create -n GPTSoVits python=3.10
conda activate GPTSoVits

pip install -r extra-req.txt --no-deps
pip install -r requirements.txt

Install FFmpeg

Conda Users

conda activate GPTSoVits
conda install ffmpeg

Ubuntu/Debian Users

sudo apt install ffmpeg
sudo apt install libsox-dev

Windows Users

Download and place ffmpeg.exe and ffprobe.exe in the GPT-SoVITS root

Install Visual Studio 2017

MacOS Users

brew install ffmpeg

Running GPT-SoVITS with Docker

Docker Image Selection

Due to rapid development in the codebase and a slower Docker image release cycle, please:

• Check Docker Hub for the latest available image tags
• Choose an appropriate image tag for your environment
• Lite means the Docker image does not include ASR models and UVR5 models. You can manually download the UVR5 models, while the program will automatically download the ASR models as needed
• The appropriate architecture image (amd64/arm64) will be automatically pulled during Docker Compose
• Docker Compose will mount all files in the current directory. Please switch to the project root directory and pull the latest code before using the Docker image
• Optionally, build the image locally using the provided Dockerfile for the most up-to-date changes

Environment Variables

• is_half: Controls whether half-precision (fp16) is enabled. Set to true if your GPU supports it to reduce memory usage.

Shared Memory Configuration

On Windows (Docker Desktop), the default shared memory size is small and may cause unexpected behavior. Increase shm_size (e.g., to 16g) in your Docker Compose file based on your available system memory.

Choosing a Service

The docker-compose.yaml defines two services:

• GPT-SoVITS-CU126 & GPT-SoVITS-CU128: Full version with all features.
• GPT-SoVITS-CU126-Lite & GPT-SoVITS-CU128-Lite: Lightweight version with reduced dependencies and functionality.

To run a specific service with Docker Compose, use:

docker compose run --service-ports <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GPT-SoVITS-CU128>

Building the Docker Image Locally

If you want to build the image yourself, use:

bash docker_build.sh --cuda <12.6|12.8> [--lite]

Accessing the Running Container (Bash Shell)

Once the container is running in the background, you can access it using:

docker exec -it <GPT-SoVITS-CU126-Lite|GPT-SoVITS-CU128-Lite|GPT-SoVITS-CU126|GPT-SoVITS-CU128> bash

Pretrained Models

If install.sh runs successfully, you may skip No.1,2,3

Users in China can download all these models here.

1. Download pretrained models from GPT-SoVITS Models and place them in GPT_SoVITS/pretrained_models.

2. Download G2PW models from G2PWModel.zip(HF)| G2PWModel.zip(ModelScope), unzip and rename to G2PWModel, and then place them in GPT_SoVITS/text.(Chinese TTS Only)

3. For UVR5 (Vocals/Accompaniment Separation & Reverberation Removal, additionally), download models from UVR5 Weights and place them in tools/uvr5/uvr5_weights.

   • If you want to use bs_roformer or mel_band_roformer models for UVR5, you can manually download the model and corresponding configuration file, and put them in tools/uvr5/uvr5_weights. Rename the model file and configuration file, ensure that the model and configuration files have the same and corresponding names except for the suffix. In addition, the model and configuration file names must include roformer in order to be recognized as models of the roformer class.

   • The suggestion is to directly specify the model type in the model name and configuration file name, such as mel_mand_roformer, bs_roformer. If not specified, the features will be compared from the configuration file to determine which type of model it is. For example, the model bs_roformer_ep_368_sdr_12.9628.ckpt and its corresponding configuration file bs_roformer_ep_368_sdr_12.9628.yaml are a pair, kim_mel_band_roformer.ckpt and kim_mel_band_roformer.yaml are also a pair.

4. For Chinese ASR (additionally), download models from Damo ASR Model, Damo VAD Model, and Damo Punc Model and place them in tools/asr/models.

5. For English or Japanese ASR (additionally), download models from Faster Whisper Large V3 and place them in tools/asr/models. Also, other models may have the similar effect with smaller disk footprint.

Dataset Format

The TTS annotation .list file format:

vocal_path|speaker_name|language|text

Language dictionary:

• ‘zh’: Chinese
• ‘ja’: Japanese
• ‘en’: English
• ‘ko’: Korean
• ‘yue’: Cantonese

Example:

D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.

Finetune and inference

Open WebUI

Integrated Package Users

Double-click go-webui.bator use go-webui.ps1 if you want to switch to V1,then double-clickgo-webui-v1.bat or use go-webui-v1.ps1

Others

python webui.py <language(optional)>

if you want to switch to V1,then

python webui.py v1 <language(optional)>

Or maunally switch version in WebUI

Finetune

Path Auto-filling is now supported

1. Fill in the audio path
2. Slice the audio into small chunks
3. Denoise(optinal)
4. ASR
5. Proofreading ASR transcriptions
6. Go to the next Tab, then finetune the model

Open Inference WebUI

Integrated Package Users

Double-click go-webui-v2.bat or use go-webui-v2.ps1 ,then open the inference webui at 1-GPT-SoVITS-TTS/1C-inference

Others

python GPT_SoVITS/inference_webui.py <language(optional)>

OR

python webui.py

then open the inference webui at 1-GPT-SoVITS-TTS/1C-inference

V2 Release Notes

New Features:

1. Support Korean and Cantonese

2. An optimized text frontend

3. Pre-trained model extended from 2k hours to 5k hours

4. Improved synthesis quality for low-quality reference audio

   more details

Use v2 from v1 environment:

1. pip install -r requirements.txt to update some packages

2. Clone the latest codes from github.

3. Download v2 pretrained models from huggingface and put them into GPT_SoVITS/pretrained_models/gsv-v2final-pretrained.

   Chinese v2 additional: G2PWModel.zip(HF)| G2PWModel.zip(ModelScope)(Download G2PW models, unzip and rename to G2PWModel, and then place them in GPT_SoVITS/text.)

V3 Release Notes

New Features:

1. The timbre similarity is higher, requiring less training data to approximate the target speaker (the timbre similarity is significantly improved using the base model directly without fine-tuning).

2. GPT model is more stable, with fewer repetitions and omissions, and it is easier to generate speech with richer emotional expression.

   more details

Use v3 from v2 environment:

1. pip install -r requirements.txt to update some packages

2. Clone the latest codes from github.

3. Download v3 pretrained models (s1v3.ckpt, s2Gv3.pth and models–nvidia–bigvgan_v2_24khz_100band_256x folder) from huggingface and put them into GPT_SoVITS/pretrained_models.

   additional: for Audio Super Resolution model, you can read how to download

V4 Release Notes

New Features:

1. Version 4 fixes the issue of metallic artifacts in Version 3 caused by non-integer multiple upsampling, and natively outputs 48k audio to prevent muffled sound (whereas Version 3 only natively outputs 24k audio). The author considers Version 4 a direct replacement for Version 3, though further testing is still needed. more details

Use v4 from v1/v2/v3 environment:

1. pip install -r requirements.txt to update some packages

2. Clone the latest codes from github.

3. Download v4 pretrained models (gsv-v4-pretrained/s2v4.pth, and gsv-v4-pretrained/vocoder.pth) from huggingface and put them into GPT_SoVITS/pretrained_models.

V2Pro Release Notes

New Features:

1. Slightly higher VRAM usage than v2, surpassing v4’s performance, with v2’s hardware cost and speed. more details

2.v1/v2 and the v2Pro series share the same characteristics, while v3/v4 have similar features. For training sets with average audio quality, v1/v2/v2Pro can deliver decent results, but v3/v4 cannot. Additionally, the synthesized tone and timebre of v3/v4 lean more toward the reference audio rather than the overall training set.

Use v2Pro from v1/v2/v3/v4 environment:

1. pip install -r requirements.txt to update some packages

2. Clone the latest codes from github.

3. Download v2Pro pretrained models (v2Pro/s2Dv2Pro.pth, v2Pro/s2Gv2Pro.pth, v2Pro/s2Dv2ProPlus.pth, v2Pro/s2Gv2ProPlus.pth, and sv/pretrained_eres2netv2w24s4ep4.ckpt) from huggingface and put them into GPT_SoVITS/pretrained_models.

Todo List

• High Priority:

  • Localization in Japanese and English.
  • User guide.
  • Japanese and English dataset fine tune training.

• Features:

  • Zero-shot voice conversion (5s) / few-shot voice conversion (1min).
  • TTS speaking speed control.
  • Enhanced TTS emotion control. Maybe use pretrained finetuned preset GPT models for better emotion.
  • Experiment with changing SoVITS token inputs to probability distribution of GPT vocabs (transformer latent).
  • Improve English and Japanese text frontend.
  • Develop tiny and larger-sized TTS models.
  • Colab scripts.
  • Try expand training dataset (2k hours -> 10k hours).
  • better sovits base model (enhanced audio quality)
  • model mix

(Additional) Method for running from the command line

Use the command line to open the WebUI for UVR5

python tools/uvr5/webui.py "<infer_device>" <is_half> <webui_port_uvr5>

This is how the audio segmentation of the dataset is done using the command line

python audio_slicer.py \
    --input_path "<path_to_original_audio_file_or_directory>" \
    --output_root "<directory_where_subdivided_audio_clips_will_be_saved>" \
    --threshold <volume_threshold> \
    --min_length <minimum_duration_of_each_subclip> \
    --min_interval <shortest_time_gap_between_adjacent_subclips>
    --hop_size <step_size_for_computing_volume_curve>

This is how dataset ASR processing is done using the command line(Only Chinese)

python tools/asr/funasr_asr.py -i <input> -o <output>

ASR processing is performed through Faster_Whisper(ASR marking except Chinese)

(No progress bars, GPU performance may cause time delays)

python ./tools/asr/fasterwhisper_asr.py -i <input> -o <output> -l <language> -p <precision>

A custom list save path is enabled

Credits

Special thanks to the following projects and contributors:

Theoretical Research

• ar-vits
• SoundStorm
• vits
• TransferTTS
• contentvec
• hifi-gan
• fish-speech
• f5-TTS
• shortcut flow matching

Pretrained Models

• Chinese Speech Pretrain
• Chinese-Roberta-WWM-Ext-Large
• BigVGAN
• eresnetv2

Text Frontend for Inference

• paddlespeech zh_normalization
• split-lang
• g2pW
• pypinyin-g2pW
• paddlespeech g2pw

WebUI Tools

• ultimatevocalremovergui
• audio-slicer
• SubFix
• FFmpeg
• gradio
• faster-whisper
• FunASR
• AP-BWE

Thankful to @Naozumi520 for providing the Cantonese training set and for the guidance on Cantonese-related knowledge.

Thanks to all contributors for their efforts

Denzii 🕊️ (@denziideng): 发现一个AI声音工具……我随便找了一段自己说话的3秒录音，丢进去试了试，结果它直接用我的声音说话的语气、自然度都像本人，吓人！

这是啥工具？

叫 CosyVoice，最大特点就是超简单声音克隆： 📍只需要 3秒 你的声音，就能完美模仿 📍支持几百种语言（普通话、英语、各种方言都行）