The team trained on 25 hours of live speech. On 525 held-out short recordings, word error rate fell from 8.81% to 5.76%, while the share of correctly written terminology rose from 43.8% to 86.7%; processing speed changed little. The authors also report more word omissions and weaker capitalization and punctuation on ordinary speech. They caution that the test labels came through the same automated transcription pipeline used for training, with selective manual review, leaving possible annotation-style overlap. The adaptation is version 0.01 and remains beta.
Glas reports fine-tuning GigaAM-v3 for Russian IT dictation
In one minute
Glas reports fine-tuning GigaAM-v3 for Russian IT dictation.
What changed: The story reports a new fine-tuning case for GigaAM-v3 in Glas, including WER and terminology gains; none of the archive documents cover this work.