Skip to content

[Not for merge] Support Zipformer encoder for LLM based ASR#1944

Draft
yfyeung wants to merge 28 commits into
k2-fsa:masterfrom
yfyeung:dev/speechllm
Draft

[Not for merge] Support Zipformer encoder for LLM based ASR#1944
yfyeung wants to merge 28 commits into
k2-fsa:masterfrom
yfyeung:dev/speechllm

Conversation

@yfyeung

@yfyeung yfyeung commented May 15, 2025

Copy link
Copy Markdown
Collaborator

This PR add zipformer_llm_zh recipe.

Some new features / modifications:

  • Introduce batch shaving mechanism, dynamically reduce batch when OOM happen to prevent training interruptions.
  • Transition to full DeepSpeed training, removing torch.autocast (supported by Fix scaling.py: ensure SwooshL/SwooshR output dtype matches input dtype #1940).
  • Expose more parameters of DynamicBucketSampler for more efficient batching.
  • Fix data preparation from huggingface.
  • Set world_size and rank explicitly for dataloader.

@yfyeung yfyeung changed the title [WIP] Add Streaming Zipformer LLM recipe for ASR [WIP] Add streaming Zipformer encoder for LLM based ASR May 15, 2025
@yfyeung yfyeung changed the title [WIP] Add streaming Zipformer encoder for LLM based ASR [WIP] Support streaming Zipformer encoder for LLM based ASR May 15, 2025
@yfyeung yfyeung changed the title [WIP] Support streaming Zipformer encoder for LLM based ASR [WIP] Support Zipformer encoder for LLM based ASR May 15, 2025
@yfyeung yfyeung changed the title [WIP] Support Zipformer encoder for LLM based ASR [Not for merge] Support Zipformer encoder for LLM based ASR Jun 2, 2025
@Johe-cqu

Johe-cqu commented Nov 2, 2025

Copy link
Copy Markdown

Wonderful work on this branch—it looks very promising. I’m exploring a Zipformer + LLM setup and found your implementation. Could you share why it hasn’t been merged into main? Also, is it currently usable (training/inference) or still experimental? Thank you for your time and contributions. @yfyeung

@yfyeung

yfyeung commented Nov 3, 2025

Copy link
Copy Markdown
Collaborator Author

Maybe you can modify the projector and remove code about blank skipping. After that, it supports training and inference, but the performance is not better than whisper encoder.

@whaozl

whaozl commented Nov 19, 2025

Copy link
Copy Markdown

Does this support streaming? Looking at the code, it doesn't seem to support streaming?

@yfyeung

yfyeung commented Nov 19, 2025

Copy link
Copy Markdown
Collaborator Author

upport streaming? Looking at the code, it doesn't seem to support streaming?

No.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants