Seed-VC

English | 简体中文 | 日本語
Currently released model supports zero-shot voice conversion 🔊 , zero-shot real-time voice conversion 🗣️ and zero-shot singing voice conversion 🎶. Without any training, it is able to clone a voice given a reference speech of 1~30 seconds.

To find a list of demos and comparisons with previous voice conversion models, please visit our demo page🌐

We are keeping on improving the model quality and adding more features.

Evaluation📊

Zero-shot voice conversion🎙🔁

We have performed a series of objective evaluations on our Seed-VC's voice conversion capabilities. For ease of reproduction, source audios are 100 random utterances from LibriTTS-test-clean, and reference audios are 12 randomly picked in-the-wild voices with unique characteristics.

Source audios can be found under ./examples/libritts-test-clean
Reference audios can be found under ./examples/reference

We evaluate the conversion results in terms of speaker embedding cosine similarity (SECS), word error rate (WER) and character error rate (CER) and compared our results with two strong open sourced baselines, namely OpenVoice and CosyVoice.
Results in the table below shows that our Seed-VC model significantly outperforms the baseline models in both intelligibility and speaker similarity.

We have also compared with non-zero-shot voice conversion models for several speakers (based on model availability):

Results show that, despite not being trained on the target speakers, Seed-VC is able to achieve significantly better results than the non-zero-shot models. However, this may vary a lot depending on the SoVITS model quality. PR or Issue is welcomed if you find this comparison unfair or inaccurate.
(Tokai Teio model from zomehwh/sovits-tannhauser)
(Matikane Tannhuaser model from zomehwh/sovits-tannhauser)
(Milky Green model from sparanoid/milky-green-sovits-4)

English ASR result computed by facebook/hubert-large-ls960-ft model
Speaker embedding computed by resemblyzer model

You can reproduce the evaluation by running eval.py script.

python eval.py 
--source ./examples/libritts-test-clean
--target ./examples/reference
--output ./examples/eval/converted
--diffusion-steps 25
--length-adjust 1.0
--inference-cfg-rate 0.7
--xvector-extractor "resemblyzer"
--baseline ""  # fill in openvoice or cosyvoice to compute baseline result
--max-samples 100  # max source utterances to go through

Before that, make sure you have openvoice and cosyvoice repo correctly installed on ../OpenVoice/ and ../CosyVoice/ if you would like to run baseline evaluation.

Zero-shot singing voice conversion🎤🎶

Additional singing voice conversion evaluation is done on M4Singer dataset, with 4 target speakers whose audio data is available here.
Speaker similariy is calculated by averaging the cosine similarities between conversion result and all available samples in respective character dataset.
For each character, one random utterance is chosen as the prompt for zero-shot inference. For comparison, we trained respective RVCv2-f0-48k model for each character as baseline.
100 random utterances for each singer type are used as source audio.

Despite Seed-VC is not trained on the target speakers, and only one random utterance is used as prompt, it still constantly outperforms speaker-specific RVCv2 models in terms of speaker similarity (SECS) and intelligibility (CER), which demonstrates the superior voice cloning capability and robustness of Seed-VC.

However, it is observed that Seed-VC's audio quality (DNSMOS) is slightly lower than RVCv2. We take this drawback seriously and will give high priority to improve the audio quality in the future.
PR or issue is welcomed if you find this comparison unfair or inaccurate.

Chinese ASR result computed by SenseVoiceSmall
Speaker embedding computed by resemblyzer model
We set +12 semitones pitch shift for male-to-female conversion and -12 semitones for female-to-male converison, otherwise 0 pitch shift

Installation📥

Suggested python 3.10 on Windows or Linux.

pip install -r requirements.txt

Usage🛠️

Checkpoints of the latest model release will be downloaded automatically when first run inference.

Command line inference:

python inference.py --source <source-wav>
--target <referene-wav>
--output <output-dir>
--diffusion-steps 25 # recommended 50~100 for singingvoice conversion
--length-adjust 1.0
--inference-cfg-rate 0.7
--f0-condition False # set to True for singing voice conversion
--auto-f0-adjust False # set to True to auto adjust source pitch to target pitch level, normally not used in singing voice conversion
--semi-tone-shift 0 # pitch shift in semitones for singing voice conversion

where:

• source is the path to the speech file to convert to reference voice
• target is the path to the speech file as voice reference
• output is the path to the output directory
• diffusion-steps is the number of diffusion steps to use, default is 25, use 50-100 for best quality, use 4-10 for fastest inference
• length-adjust is the length adjustment factor, default is 1.0, set <1.0 for speed-up speech, >1.0 for slow-down speech
• inference-cfg-rate has subtle difference in the output, default is 0.7
• f0-condition is the flag to condition the pitch of the output to the pitch of the source audio, default is False, set to True for singing voice conversion
• auto-f0-adjust is the flag to auto adjust source pitch to target pitch level, default is False, normally not used in singing voice conversion
• semi-tone-shift is the pitch shift in semitones for singing voice conversion, default is 0

Gradio web interface:

python app.py

Then open the browser and go to http://localhost:7860/ to use the web interface.

Real-time voice conversion GUI:

python real-time-gui.py

IMPORTANT: It is strongly recommended to use a GPU for real-time voice conversion.
Some performance testing has been done on a NVIDIA RTX 3060 Laptop GPU, results and recommended parameter settings are listed below:

You can adjust the parameters in the GUI according to your own device performance, the voice conversion stream should work well as long as Inference Time is less than Block Time.
Note that inference speed may drop if you are running other GPU intensive tasks (e.g. gaming, watching videos)
Generally, latency is around 1~2s to prevent quality drop (the sad nature of diffusion models...😥), but we are keeping on looking for ways to reduce it.

(GUI and audio chunking logic are modified from RVC, thanks for their brilliant implementation!)

TODO📝

• [x] Release code
• [x] Release v0.1 pretrained model:
• [x] Huggingface space demo:
• [x] HTML demo page (maybe with comparisons to other VC models): Demo
• [x] Streaming inference
• [ ] Reduce streaming inference latency
• [ ] Demo video for real-time voice conversion
• [x] Singing voice conversion
• [ ] Noise resiliency for source & reference audio
  • [x] Source audio is noise resilience
• [ ] Potential architecture improvements
  • [x] U-ViT style skip connections
  • [x] Changed input to OpenAI Whisper
• [ ] Code for training on custom data
• [x] Changed to BigVGAN from NVIDIA for singing voice decoding
• [x] Whisper version model for singing voice conversion
• [x] Objective evaluation and comparison with RVC/SoVITS for singing voice conversion
• [x] Improve audio quality
• [ ] More to be added

CHANGELOGS🗒️

• 2024-10-28:
  • Updated fine-tuned 44k singing voice conversion model with better audio quality
• 2024-10-27:
  • Added real-time voice conversion GUI
• 2024-10-25:
  • Added exhaustive evaluation results and comparisons with RVCv2 for singing voice conversion
• 2024-10-24:
  • Updated 44kHz singing voice conversion model, with OpenAI Whisper as speech content input
• 2024-10-07:
  • Updated v0.3 pretrained model, changed speech content encoder to OpenAI Whisper
  • Added objective evaluation results for v0.3 pretrained model
• 2024-09-22:
  • Updated singing voice conversion model to use BigVGAN from NVIDIA, providing large improvement to high-pitched singing voices
  • Support chunking and streaming output for long audio files in Web UI
• 2024-09-18:
  • Updated f0 conditioned model for singing voice conversion
• 2024-09-14:
  • Updated v0.2 pretrained model, with smaller size and less diffusion steps to achieve same quality, and additional ability to control prosody preservation
  • Added command line inference script
  • Added installation and usage instructions