MediumVC is an utterance-level method towards any-to-any VC. Before that, we propose SingleVC to perform A2O tasks(X_i → Ŷ_i) , X_i means utterance i spoken by X). The Ŷ_i are considered as SSIF. To build SingleVC, we employ a novel data augment strategy: pitch-shifted and duration-remained(PSDR) to produce paired asymmetrical training data. Then, based on pre-trained SingleVC, MediumVC performs an asymmetrical reconstruction task(Ŷ_i → X̂_i). Due to the asymmetrical reconstruction mode, MediumVC achieves more efficient feature decoupling and fusion. Experiments demonstrate MediumVC performs strong robustness for unseen speakers across multiple public datasets. Here is the official implementation of the paper, MediumVC.

The following are the overall model architecture.

For the audio samples, please refer to our demo page. The more converted speeches can be found in "Demo/ConvertedSpeeches/".

You can install the dependencies with

pip install -r requirements.txt

Dvector is a robust speaker verification (SV) system pre-trained on VoxCeleb1 using GE2E loss, and it produces 256-dim speaker embedding. In our evaluation on multiple datasets(VCTK with 30000 pairs, Librispeech with 30000 pairs and VCC2020 with 10000 pairs), the equal error rates(EERs)and thresholds(THRs) are recorded in Table. Then Dvector with THRs is also employed to calculate SV accuracy(ACC) of pairs produced by MediumVC and other contrast methods for objective evaluation. The more details can access paper.

The HiFi-GAN vocoder is employed to convert log mel-spectrograms to waveforms. The model is trained on universal datasets with 13.93M parameters. Through our evaluation, it can synthesize 22.05 kHz high-fidelity speeches over 4.0 MOS, even in cross-language or noisy environments.

You can download the pretrained model, and then edit "Any2Any/infer/infer_config.yaml".Test Samples could be organized as "wav22050/$figure$/*.wav".

python Any2Any/infer/infer.py

The corpus should be organized as "VCTK22050/$figure$/*.wav", and then edit the config file "Any2Any/pre_feature/preprocess_config.yaml".The output "spk_emb_mel_label.pkl" will be used for training.

python Any2Any/pre_feature/figure_spkemb_mel.py

Please edit the paths of pretrained hifi-model,wav2mel,dvector,SingleVC in config file "Any2Any/config.yaml" at first.

python Any2Any/solver.py