AI News/Wire summary
CAFE confines feature distillation to adapters to improve compact multilingual speech recognition
Key points
- Study finds feature distillation does not add to logit distillation for two common feature losses in compact multilingual ASR.
- CKA loss lowers WER on non-English languages but hurts data-rich English, per the paper's gradient analysis.
- CAFE routes the CKA gradient into adapters English bypasses, cutting average WER from 8.55% to 8.02% on Open ASR Leaderboard.
- At 254M parameters, CAFE trails its 0.6B teacher by 2 WER points across 48 test sets in 24 languages.
Why it matters for marketers: Teams distilling compact multilingual speech models may get non-English gains without English regressions by isolating feature losses in language-specific adapters.
Summary by the Z Media news desk, from the outlets' reporting. Read the full stories at the source.
Read the original
Related on Z Media
- Anthropic gives METR researchers employee-level access to its modelsvia TechCrunch, 5 hr ago
- Microsoft releases open-source tool to improve two-to-six-week weather forecastsvia The Verge, yesterday
- AI solves far more theory problems across several physics and math fieldsvia Wired, 2 days ago
- Bringing GPT-Rosalind to academic research labsvia OpenAI, 6 hr ago
- Google Antigravity helps scientists find a candidate antimicrobial peptidevia Google DeepMind, 9 hr ago
- Google's Gemini 4 Argon is now open to everyone, with mixed early reviewsvia Ben's Bites, 12 hr ago