Congratulations!! That’s wonderful!! 🎉🍾

Congrats!!! 🎉

CTC-DRO can be applied to ASR with minimal computational costs, and offers the potential for reducing group disparities in other domains with similar challenges.

📄 Read our paper: arxiv.org/pdf/2502.017...
💻 Get the code: github.com/Bartelds/ctc...

The result:
📊 Worst-language error ↓ up to 47.1%
📊 Average error ↓ up to 32.9%

CTC-DRO works seamlessly with existing self-supervised speech models through ESPnet 🚀

We present CTC-DRO, which addresses the shortcomings of the group DRO objective by:
✅ Input length-matched batching to mitigate CTC’s scaling issues
✅ Smoothing the group weight update to prevent overemphasis on consistently high-loss groups

Why? Group DRO needs comparable training losses between languages. But in ASR, CTC-based losses vary due to differences in speech length, speakers, and acoustics. This creates spurious differences across language groups.

Result? Worse performance.

We need a new approach 🚀

CTC-based fine-tuning has been successful in multilingual ASR benchmarks but it doesn't fix language performance gaps. Group DRO could help by focusing on worst-performing languages, but it does not work ❌

🙋‍♂️

Mentioning this post from @cjziems.bsky.social, listing some starter packs: bsky.app/profile/cjzi...

🙋‍♂️

👋