bshall / soft-vc

Soft speech units for voice conversion

Soft Speech Units for Improved Voice Conversion

Official repository for A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion. Audio samples can be found here. Colab demo can be found here.

Abstract: The goal of voice conversion is to transform source speech into a target voice, keeping the content unchanged. In this paper, we focus on self-supervised representation learning for voice conversion. Specifically, we compare discrete and soft speech units as input features. We find that discrete representations effectively remove speaker information but discard some linguistic content – leading to mispronunciations. As a solution, we propose soft speech units learned by predicting a distribution over the discrete units. By modeling uncertainty, soft units capture more content information, improving the intelligibility and naturalness of converted speech.

For modularity, each component of the system is housed in a separate repository. Please visit the following links for more details:

• HuBERT content encoders
• Acoustic Models
• HiFiGAN vocoder

^{Fig 1: Architecture of the voice conversion system. a) The discrete content encoder clusters audio features to produce a sequence of discrete speech units. b) The soft content encoder is trained to predict the discrete units. The acoustic model transforms the discrete/soft speech units into a target spectrogram. The vocoder converts the spectrogram into an audio waveform.}

Example Usage

Programmatic Usage

import torch, torchaudio

# Load the content encoder (either hubert_soft or hubert_discrete)
hubert = torch.hub.load("bshall/hubert:main", "hubert_soft", trust_repo=True).cuda()

# Load the acoustic model (either hubert_soft or hubert_discrete)
acoustic = torch.hub.load("bshall/acoustic-model:main", "hubert_soft", trust_repo=True).cuda()

# Load the vocoder (either hifigan_hubert_soft or hifigan_hubert_discrete)
hifigan = torch.hub.load("bshall/hifigan:main", "hifigan_hubert_soft", trust_repo=True).cuda()

# Load the source audio
source, sr = torchaudio.load("path/to/wav")
assert sr == 16000
source = source.unsqueeze(0).cuda()

# Convert to the target speaker
with torch.inference_mode():
    # Extract speech units
    units = hubert.units(source)
    # Generate target spectrogram
    mel = acoustic.generate(units).transpose(1, 2)
    # Generate audio waveform
    target = hifigan(mel)

Citation

If you found this work helpful please consider citing our paper:

@inproceedings{
    soft-vc-2022,
    author={van Niekerk, Benjamin and Carbonneau, Marc-André and Zaïdi, Julian and Baas, Matthew and Seuté, Hugo and Kamper, Herman},
    booktitle={ICASSP}, 
    title={A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion}, 
    year={2022}
}

soft-vc's Issues

Interesting work!

Hi Benjamin,

Looked at your demo page, it looks nice!
Will the paper be on arxiv?
Looking forward to it.

Can you share the training scripts?

I have tried the inference example and the result is exciting.

Can you share the training scripts to me or open source?

This will help me a lot.

Thanks!

Difference between SSL and PPG-based methods?

Hi, I really appreciate your work; the demo sounds great.
I also read papers about PPG-based VC, which uses ASR for PPG extraction. I just wonder about the difference between SSL and PPG-based methods. It seems they both extract some information about linguistics. Have you ever compared them?
Thank you!

How to get discrete units from soft content encoder as its predicting discrete tokens in theory but in implementation its otherwise?

Bug：TypeError: hubert_soft() got an unexpected keyword argument 'trust_repo'

I would like to ask if anyone has encountered this problem while doing experiments? How was it solved?

Using cache found in /data0/home/Liqy/.cache/torch/hub/bshall_hubert_main
Traceback (most recent call last):
File "/tmp/pycharm_project_363/test.py", line 4, in
hubert = torch.hub.load("bshall/hubert:main", "hubert_soft", trust_repo=True).cuda()
File "/Nas/Liqy/hubert-main/lib/python3.10/site-packages/torch/hub.py", line 404, in load
model = _load_local(repo_or_dir, model, *args, **kwargs)
File "/Nas/Liqy/hubert-main/lib/python3.10/site-packages/torch/hub.py", line 433, in _load_local
model = entry(*args, **kwargs)
TypeError: hubert_soft() got an unexpected keyword argument 'trust_repo'

Is it my pytorch version? Can you give me your pytorch version for reference? Thank you very much

speech resynthesis?

Hi, thanks for the sample codes! very easy to use with impressive results! I am wondering if it is possible to resynthesize the speaker's voice, instead of speech conversion, using your model?

can the pre-trained hubert-soft or discrete model be used for encoding mandarin Chinese language data ?

Hi @bshall , can the pre-trained hubert-soft or discrete model be used for encoding mandarin Chinese language data ? I want to train a hubert soft model for Chinese language VC.

skipped phonemes in generated audio

hi, thank you for sharing your code.

i am trying to do voice conversion from English speech to Vietnamese speaker. to do that, i did the following steps

• extract units for both English and Vietnamese dataset
• train kmeans on both types of units & extract discrete labels
• train soft encoder
• extract soft units
• train acoustic model
• train hifigan on Vietnamese dataset

the output for Vietnamese speech (input audio is Vietnamese, of a different speaker) is okay. but output for English is not that good. phonemes are often skipped or mispronouced. do you have any suggestions on how i can improve the results?

is real-time voice conversion possible?

Hi- very impressed by the VC framework. It's very fast and accurate.
I'm wondering is real-time possible? I have a simple WS server that receives audio, but when i push the data through soft-vc, the end result is just noise. In the code below, I save the input stream just to confirm the audio is being received correctly (which it is).
Here is a snippet of my code:

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
hubert = torch.hub.load("bshall/hubert:main", "hubert_soft").cuda()

acoustic_load_path = "./pretrained_models/acoustic.pt"
checkpoint = torch.load(acoustic_load_path, map_location=device)["acoustic-model"]
acoustic = AcousticModel().to(device)
acoustic.load_state_dict(checkpoint)
acoustic.eval()

# load custom vocoder
hifigan_load_path = "./pretrained_models/hifigan.pt"
checkpoint = torch.load(hifigan_load_path, map_location=device)[
    "generator"]["model"]
hifigan = HifiganGenerator().to(device)
consume_prefix_in_state_dict_if_present(checkpoint, "module.")
hifigan.load_state_dict(checkpoint)
hifigan.eval()
hifigan.remove_weight_norm()



inputs = []
outputs = []
while True:
    data = None
    try:
        data = await websocket.recv()
    except:
        break
    if isinstance(data, str):
        print(f"string -> {data}")
        continue

    source = torch.from_numpy(numpy.frombuffer(
        data, dtype=numpy.int16).astype('float32') / 32767)
    source = source.reshape((1, -1))
    source = source.unsqueeze(0).cuda()
    # # Convert to the target speaker
    with torch.inference_mode():
        # Extract speech units
        units = hubert.units(source)
        # Generate target spectrogram
        mel = acoustic.generate(units).transpose(1, 2)
        # Generate audio waveform
        target = hifigan(mel)
        inputs.append(source.squeeze(0).cpu())
        outputs.append(target.squeeze(0).cpu())
        await ws.send(data)

print(f"saving files...")
input_result = torch.cat(inputs, dim=1)
torchaudio.save("inputs.wav", input_result, sample_rate=16_000)
output_result = torch.cat(outputs, dim=1)
torchaudio.save("outputs.wav", output_result, sample_rate=16_000)

K-means training

What accuracy did you achieve while training the k-means model in the content encoder?

Discrete content encoder example

@bshall since this repo gives a simple inference example based on a soft content encoder. Would you give a discrete encoder based inference example? I found some issues when I try to use a discrete encoder. Here is my code:

when it run to mel = acoustic.generate(units).transpose(1, 2)
a dimension mismatch occurred. Looking forward to your suggestions, thanks a lot.

File "D:\Software-location\Anaconda\lib\site-packages\torch\autograd\grad_mode.py", line 27, in decorate_context
return func(*args, **kwargs)
File "C:\Users\YoungTown/.cache\torch\hub\bshall_acoustic-model_main\acoustic\model.py", line 23, in generate
x = self.encoder(x)
File "D:\Software-location\Anaconda\lib\site-packages\torch\nn\modules\module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "C:\Users\YoungTown/.cache\torch\hub\bshall_acoustic-model_main\acoustic\model.py", line 49, in forward
x = self.convs(x.transpose(1, 2))
IndexError: Dimension out of range (expected to be in range of [-2, 1], but got 2)

import torch, torchaudio
hubert = torch.hub.load("bshall/hubert:main", "hubert_discrete").cuda()
acoustic = torch.hub.load("bshall/acoustic-model:main", "hubert_discrete").cuda()
hifigan = torch.hub.load("bshall/hifigan:main", "hifigan_hubert_discrete").cuda()

source, sr = torchaudio.load("./huohuo.wav")
source = torchaudio.transforms.Resample(orig_freq=sr, new_freq=16000)(source)
source = source[0].unsqueeze(0).unsqueeze(0).cuda()

with torch.inference_mode():
    units = hubert.units(source)
    mel = acoustic.generate(units).transpose(1, 2)
    target = hifigan(mel)

    target = target.squeeze().cpu()
    target = target.unsqueeze(0).cpu()
    torchaudio.save('./huohuo_new.wav', target, 16000)

About fine-tuning.

Hi @bshall . I got impressive results trying to convert singing voice samples. So, I was trying to understand how to fine-tune a specific singer's voice. Do I need to train each individual component for this?

inference and training scripts

Hello, do You plan to add local training and inference scripts other than jupyter notebook demo? thanks in advance!

Will the training code be available?

Nice work, I found that there is only the inference code here, will the training code be also released?

Found it!