Observed facts: Glas fine-tuned GigaAM-v3 on 25.1 hours of live speech. On 525 held-out short recordings, WER fell from 8.81% to 5.76%, and correctly written terminology rose from 43.8% to 86.7%. The authors say processing speed changed little, but word omissions increased and capitalization and punctuation worsened on ordinary speech.
Mechanism: The team retained the deployed RNNT model architecture and focused adaptation on upper encoder layers and the RNNT head, using a terminology-aware dataset and explicit transcription conventions. The optimization target was a domain-specific failure mode—English names, loanwords, and abbreviations—not speech recognition in the abstract. This is why terminology accuracy can rise while other output qualities regress.
Why now: The case gives builders a concrete example of a relatively bounded adaptation path for a production speech interface, alongside metrics that make trade-offs visible. First implication: A specialized recognizer should be evaluated against both domain terminology and ordinary-language quality, rather than a single aggregate score.
Evidence boundary: The reported test set contains 525 short recordings and was labeled through the same automated pipeline used for training, with selective manual review. The authors explicitly note that annotation-style overlap may account for part of the measured gain. The result is version 0.01 beta, not evidence of broad independent replication.
First watch condition: Independent evaluation with human-verified labels should establish whether terminology gains persist without increased omissions or regressions in punctuation and capitalization.