Martijn Bartelds
@mbartelds.bsky.social
Postdoctoral Scholar Stanford NLP
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
✅ Input length-matched batching to mitigate CTC’s scaling issues
✅ Smoothing the group weight update to prevent overemphasis on consistently high-loss groups
March 12, 2025 at 3:29 PM
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
✅ 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 🚀
Result? Worse performance.
We need a new approach 🚀
March 12, 2025 at 3:29 PM
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 🚀
Result? Worse performance.
We need a new approach 🚀
🎙️ Speech recognition is great - if you speak the right language.
Our new @stanfordnlp.bsky.social paper introduces CTC-DRO, a training method that reduces worst-language errors by up to 47.1%.
Work w/ Ananjan, Moussa, @jurafsky.bsky.social, Tatsu Hashimoto and Karen Livescu.
Here’s how it works 🧵
Our new @stanfordnlp.bsky.social paper introduces CTC-DRO, a training method that reduces worst-language errors by up to 47.1%.
Work w/ Ananjan, Moussa, @jurafsky.bsky.social, Tatsu Hashimoto and Karen Livescu.
Here’s how it works 🧵
March 12, 2025 at 3:29 PM
🎙️ Speech recognition is great - if you speak the right language.
Our new @stanfordnlp.bsky.social paper introduces CTC-DRO, a training method that reduces worst-language errors by up to 47.1%.
Work w/ Ananjan, Moussa, @jurafsky.bsky.social, Tatsu Hashimoto and Karen Livescu.
Here’s how it works 🧵
Our new @stanfordnlp.bsky.social paper introduces CTC-DRO, a training method that reduces worst-language errors by up to 47.1%.
Work w/ Ananjan, Moussa, @jurafsky.bsky.social, Tatsu Hashimoto and Karen Livescu.
Here’s how it works 🧵