[WIP] Add a SSL recipe on LibriSpeech dataset

[yfyeung]

• Gradient accumulation mechanism
• HuBERT fine-tune recipe based on transformers, which supports the pre-trained checkpoints from:
  - huggingface/facebook: hubert-base-ls960, hubert-large-ll60k, hubert-xlarge-ll60k
  - huggingface/TencentGameMate: chinese-hubert-base, chinese-hubert-large
• HuBERT pre-train recipe based on transformers
• Release with the paper

[yfyeung]

The HuBERT finetuning recipe with the TencentGameMate chinese-hubert-large was utilized in track 1 of the ICASSP2024 ICMC-ASR Grand Challenge, achieving a CER of 18.18 directly in the evaluation.

[yfyeung]

We will soon release the Zipformer pre-training and fine-tuning guide alongside the paper.
We provide recipes for both Transformer-based and Zipformer-based HuBERT.
Additionally, checkpoints from fairseq and transformers are compatible.
Our training framework achieves faster convergence, better performance, and is memory efficient and stable at half-precision.

[TueVNguyen]

@yfyeung Could you provide me with the datasets from the ICASSP 2024 ICMC-ASR Grand Challenge? Unfortunately, the challenge has already concluded, and I am unable to register to experiment with it.

[yfyeung]

@yfyeung Could you provide me with the datasets from the ICASSP 2024 ICMC-ASR Grand Challenge? Unfortunately, the challenge has already concluded, and I am unable to register to experiment with it.

Sure
https://pan.baidu.com/s/1Sc4nJ0a4FzFHBwoOpyGpUQ?pwd=6y7e

[yfyeung]

See: #1500