Gemma Embedding 2 hits SOTA on Vietnamese speech after 3,000 hours of audio, 2 hours on 8xH100
stablequan · x · 2026-10-14
Developer stablequan reports that fine-tuning Google's Gemma Embedding 2 on 3,000 hours of Vietnamese audio produced a SOTA speech embedding model — and the entire training run took just 2 hours on 8xH100.
A striking data point that for mid- and low-resource languages, fine-tuning an open embedding backbone on local data can hit top results at tiny cost.
More from Models
- Why three frontier models with different training beats one: doodlestein's case for Argon as the third mind — doodlestein · 2026-10-14
- Researcher calls out Claude's 'Memory updated' banner: a warning nobody ever needs — adam_dorr · 2026-10-14
- Google Argon emails surface: AI Pro subscribers getting early access to the new frontier model — raghavrdx · 2026-10-14
- DeepMind Details Gemini 4 Argon's CoT Monitor That Intervenes When Reasoning Goes Off the Rails — danielrock · 2026-10-14
- Investor bindureddy: DeepSeek Flash is the best nearly-free coding model for the masses — bindureddy · 2026-10-14
- Sonnet 4 Retires in Two Days as Users Bid Farewell to Anthropic's Steadiest Model — repligate · 2026-10-14