Open-source ASR for low-resource languages: fine-tuned Whisper hits 3.2% WER on Latvian
Slight_Republic_4242 · reddit · 2026-09-14
University of Latvia researchers released LATE, an open-source ASR toolkit for low-resource languages with local/cloud deployment, a statically compiled backend, and private inference. Fine-tuning Whisper Large V3 on 273.3 hours of Latvian speech cut WER from 19.2% to 3.2% on Common Voice and from 29.1% to 12.8% on a harder media set. For Latgalian (150k speakers) with just 40.2 hours of data, transfer learning from the Latvian model achieved 9.1% WER. Quantized models are included for constrained hardware.
More from coding & agent
- foliate-js: The 1.1k-Star Library Powering an Obsidian E-Book Reader — vista8 · 2026-09-14
- Matt Pocock: Every Org Is Desperate to Adopt AI Coding — Become the Person Who Delivers It — mattpocockuk · 2026-09-14
- Red-teaming public-facing AI agents: quick wins to make your agent safer — njyx · 2026-09-14
- What Should a Coding Agent Be Allowed to Do? .env Key Exposure Sparks Debate — radim11 · 2026-09-14
- South African founders run immersive Vibe Coding workshops for business owners — huangyun_122 · 2026-09-14
- Training Agents series wraps with RL environments for training coding agents — SergioPaniego · 2026-09-14