k2SSL: A Faster and Better Framework for Self-Supervised Speech Representation Learning

[yfyeung]

• Libri-Light data processing script
• Libri-Light Zipformer multi-node-multi-gpu pre-train recipe
• LibriSpeech Zipformer bpe-level prund rnn-t fine-tune recipe
• LibriSpeech Zipformer letter-level ctc fine-tune recipe
• Release all resource and results for Zipformer Base
• Release all resource and results for Zipformer Large

[yfyeung]

Part of the resources have been released:

Zipformer Base pre-trained with cross entropy loss: Checkpoints, logs, and scripts
Zipformer Base fine-tuned with pruned RNN-T loss: Checkpoints, logs, and scripts
Zipformer Base fine-tuned with letter-level CTC loss: Checkpoints, logs, and scripts
Zipformer Base pre-trained manifests including kmeans labels: Dataset

With these resources, I believe anyone with 8 V100 32G GPUs can easily reproduce our experiments.

[ezerhouni]

@yfyeung Thank you very much for the PR and sharing the model weights. Do you plan to release a paper also ?

[yfyeung]

@yfyeung Thank you very much for the PR and sharing the model weights. Do you plan to release a paper also ?

Maybe, but currently I am not sure whether it is suitable for a technical report or just a normal research paper.