Skip to content

SOTA method for self-supervised speaker verification leveraging a large-scale pretrained ASR model.

License

Notifications You must be signed in to change notification settings

theolepage/wavlm_ssl_sv

Repository files navigation

wavlm_ssl_sv

This repository contains the source code of the article Towards Supervised Performance on Speaker Verification with Self-Supervised Learning by Leveraging Large-Scale ASR Models (INTERSPEECH 2024) [arXiv].

The proposed framework fine-tunes a pre-trained WavLM using pseudo-labels, generated through Self-Supervised Learning (SSL), for Speaker Verification (SV). Initial pseudo-labels are derived from an SSL DINO-based model and are iteratively refined by clustering the model embeddings.

Our method achieves 0.99% EER on VoxCeleb1-O, establishing the new SOTA on Speaker Verification with SSL.

Please refer to the article for more details on the implementation and a comparative study with other works.


Usage

Installation

  • Install dependencies with pip install -r requirements.txt.
  • Prepare data for VoxCeleb, MUSAN, and RIR datasets following voxceleb_trainer.
  • Download WavLM-Base+ model and place WavLM-Base+.pt at the root folder.

Training

Step 1: Extract DINO speaker embeddings

The code to train the DINO model is not currently provided. We recommend using sslsv or 3D-Speaker to extract initial speaker embeddings.

Alternatively, you can directly download the DINO embeddings we used for our system: dino_vox2_embeddings.pt.

Note: the embeddings file must be a Dict[str, torch.Tensor] representing all VoxCeleb2 samples with the following format for keys: id00012/21Uxsk56VDQ/00001.wav.

Step 2: Generate pseudo-labels

python pseudo_labeling.py PATH_TO_EMBEDDINGS_FILE PATH_TO_PL_FILE

Step 3: Fine-tune WavLM MHFA

python trainSpeakerNet.py --config configs/wavlm_mhfa_dlg_lc.yaml --train_list PATH_TO_PL_FILE --distributed

Iterative process

  1. Extract embeddings from the WavLM MHFA model: python trainSpeakerNet_Eval.py --config configs/wavlm_mhfa_dlg_lc.yaml --generate_embeddings --embeddings_path PATH_TO_EMBEDDINGS_FILE.

  2. Repeat steps 2 and 3. Make sure to change save_path in the config to avoid overwriting the existing model.

Step 4: Large-Margin Fine-Tuning

  1. Copy the latest model checkpoint to exp/wavlm_mhfa_dlg_lc_lmft/model to resume training.

  2. Start training: python trainSpeakerNet.py --config configs/wavlm_mhfa_dlg_lc_lmft.yaml --train_list PATH_TO_PL_FILE --distributed.

Evaluation

python trainSpeakerNet_Eval.py --config configs/wavlm_mhfa_dlg_lc_lmft.yaml --eval

Model weights

The checkpoint of our best model reaching 0.99% EER on VoxCeleb1-O is available for download: wavlm_mhfa_dlg_lc_lmft.


Acknowledgements

This repository contains third-party components and code adapted from other open-source projects, including: SLT22_MultiHead-Factorized-Attentive-Pooling and Loss-Gated-Learning.


Citation

If you use this project, please consider starring this repository on GitHub and citing the following paper.

@InProceedings{miara2024WavLMSSLSV,
  author    = {Miara, Victor and Lepage, Théo and Dehak, Réda},
  booktitle = {INTERSPEECH},
  title     = {Towards Supervised Performance on Speaker Verification with Self-Supervised Learning by Leveraging Large-Scale ASR Models},
  year      = {2024},
  url       = {https://arxiv.org/abs/2406.02285},
}

License

This project is released under the MIT License.