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.

Original post →

More from coding & agent

coding & agent channel →