voice conversion system
This repository provides a PyTorch implementation of AutoVC-WavRNN
The audio demo for AUTOVC-WavRNN can be found in results.
To get the audio:
1.Load and rescale the wav with the max absolute value.
2.Normalize the volume of wavs.(target dBFS is -30)
3.Skip utterances that are too short.(less than 1.5s)
To get the mel-spectrogram:
4.Preemphasis.(filter coefficient is 0.97)
5.STFT.(n_fft=1024, hop_size=256,win_size=1024)
6.Build mel filter bank.
7.Inner product the result of STFT and mel filter bank then get 80 channel mel-spectrogram.
8.Transform amplitude to dB.(ref_level_db=16)
9.Normalize mel-spectrogram to [0,1].
try:
python synthesizer_preprocess_audio.py /datato preprocess audio
try:
python synthesizer_preprocess_embeds.py /data/SV2TTS_autovc-ttsdb/synthesizer_vad -e pretrained.ptto preprocess speaker embedding
speaker encoder pretrained model: https://drive.google.com/file/d/1n1sPXvT34yXFLT47QZA6FIRGrwMeSsZc/view.
the autoVC-WavRNN model which trained by mixed VCTK dataset and VCC2020 dataset is in the folder of
/model, please choose all rar files and decompression.
the AutoVC training step is included in the
autovc_train.py, the model of AutoVC included in themodel_vc.py.
try:
python autovc_train.py autovc-vcc2020 /data -g
try:
python vocoder_train.py my_vocoder /data/ -g
try:
python convert.py
CorentinJ/Real-Time-Voice-Cloning: https://github.com/CorentinJ/Real-Time-Voice-Cloning
auspicious3000/autovc: https://github.com/auspicious3000/autovc
non-parallel many-to-many voice conversion, as well as zero-shot voice conversion, remain under- explored areas. Deep style transfer algorithms, such as generative adversarial networks (GAN) and conditional variational autoencoder (CVAE), are being applied as new solutions in this field. However, GAN training is sophisticated and diffi- cult, and there is no strong evidence that its gen- erated speech is of good perceptual quality.. On the other hand, CVAE training is simple but does not come with the distribution-matching property of a GAN. In this paper, we propose a new style transfer scheme that involves only an autoencoder with a carefully designed bottleneck. We formally show that this scheme can achieve distribution- matching style transfer by training only on a self- reconstruction loss. Based on this scheme, we proposed AUTOVC, which achieves state-of-the- art results in many-to many voice conversion with non-parallel data, and which is the first to perform zero-shot voice conversion. First, most voice conversion systems assume the availability of parallel training data speech pairs where the two speakers utter the same sentences. Only a few can be trained on non-parallel data. Second, among the few existing algorithms that work on non-parallel data, even fewer can work for many-to-many conversion, i.e. converting from multiple source speakers to multiple target speakers. Last but not least, no voice conver- sion systems are able to perform zero-shot conversion, i.e. conversion to the voice of an unseen speaker by looking at only a few of his/her utterances no change was made. https://1drv.ms/v/s!AjJEvYoU3ZQOeaX5sBlNrmBUwco https://1drv.ms/b/s!AjJEvYoU3ZQOei_ja4cqd5N-KnQ https://github.com/freenowill/AutoVC-WavRNN https://1drv.ms/b/s!AjJEvYoU3ZQOgQDRna1IVFAUF0BP https://1drv.ms/b/s!AjJEvYoU3ZQOf5ZkDIgDyc4cEVQ https://1drv.ms/b/s!AjJEvYoU3ZQOgQLaTxPruXGcOl29 https://1drv.ms/b/s!AjJEvYoU3ZQOfFdd_SrfiTWTXrM https://1drv.ms/b/s!AjJEvYoU3ZQOfVpOd_Vl3iKkXxI https://1drv.ms/b/s!AjJEvYoU3ZQOgQMBd30zqSmJsll7 Hadis Mahmoudi Bardzard, a master's student in medical engineering, majoring in bioelectricity. Project number 31 Titled zero-shot voice style conversion with only auto-encoder loss hadismahmoudi21@gmail.com https://1drv.ms/b/s!AjJEvYoU3ZQOe-zUlwbufIhIM9k https://1drv.ms/v/s!AjJEvYoU3ZQOdeVebl3jjuVU9O4 https://1drv.ms/v/s!AjJEvYoU3ZQOdgIJcKN6c4z7lBE https://1drv.ms/v/s!AjJEvYoU3ZQOeaX5sBlNrmBUwco