Glas fine-tunes GigaAM-v3 speech recognition: IT terms correct 87% of the time instead of 44%
Dictation app Glas, built on the GigaAM-v3 speech recognition model, wrote Docker as a Russian phonetic spelling ("dokker"), and turned the Russian sentence "My AI rarely makes mistakes" into "My rarely makes mistakes" (the abbreviation for "AI" got lost). After fine-tuning on 25 hours of live speech, the share of correctly written IT terms rose from 43.8% to 86.7%, and the word error rate (WER) fell from 8.81% to 5.76%. This is a beta, and the test set was labeled by the same pipeline as the training data, so check the result on your own data.
AI-processed from Хабр: NLP; edited by Hamidun News
If you dictate code reviews or work notes, you know the pain: speech recognition handles "hi, how are you" without trouble but breaks on words from a developer's daily life. The dictation app Glas runs on GigaAM-v3, a Russian speech recognition model, and IT vocabulary is exactly where it stumbled: Docker turned into a Russian phonetic spelling ("dokker"), and Python into another ("Payton"). The Glas team fine-tuned the model on 25 hours of live speech, and the share of correctly written IT terms rose from 43.8% to 86.7%, from "less than half" to "almost always". Overall word error rate (WER) fell from 8.81% to 5.76%, roughly 35% relative to the original value.
What exactly was breaking
The base model understood ordinary Russian speech well, but it failed on three groups of words: English names (Docker, Kubernetes), established loanwords such as "dashboard" in its Russian spelling, and the abbreviation for "AI" (written "II" in Russian). Sometimes whole words disappeared. The model transcribed the Russian sentence "My AI rarely makes mistakes" as "My rarely makes mistakes", dropping the abbreviation.
A simple replacement dictionary only helps partly. In tests, Kubernetes alone came out as "Kyubernit", "Kubernite", "kubernetis", "Gubernitis" and "Qubernit" (the first three in Cyrillic). You can't put every such variant in a dictionary. The abbreviation for "AI" is even trickier: in text alone it is hard to tell it from a drawn-out Russian conjunction "i" ("and"), so you need the audio. That means the model itself has to be fixed, not the post-processing.
How it was fine-tuned and what came out
GigaAM is a family of speech recognition models built on the Conformer architecture. The team uses the variant with an RNNT head, which outputs text with casing and punctuation directly. On the server, the encoder runs in FP16 and the RNNT head in FP32. Silero VAD cuts long recordings at pauses, and short segments are recognized in batches. Decoding is greedy, with no external language model, and the hardware is an RTX A2000 GPU with 12 GB.
The tokenizer did not need changing: its 1024-token vocabulary already includes Latin letters, so the needed names are encoded without workarounds. The team also kept the architecture, model code and weight format unchanged, so the fine-tuned version can replace the old one without reworking the server.
The final fine-tuning ran on 25 hours of live speech. Evaluation on 525 held-out short recordings gave these numbers:
- WER fell from 8.81% to 5.76%;
- the share of correctly written terms rose from 43.8% to 86.7%;
- speed barely changed.
For production, the last point matters as much as the first two: quality improved with no extra GPU cost.
Lesson from the first failed attempt
The team first took the CTC variant of GigaAM and assembled 4.4 hours of data: about 1100 synthetic phrases with IT terms and roughly as many recordings from the Russian and English parts of the FLEURS dataset. They froze the bottom 12 encoder layers and trained the top four layers and the CTC head.
On a test bench of 12 synthetic phrases, the share of correct terms jumped from 6% to 86%. But when compared with the production RNNT model on 150 live FLEURS recordings, the picture looked different. The takeaway for practitioners: a good result on an artificial test guarantees nothing, so you have to validate on live speech.
Honest limitations
The authors call the release version 0.01 and a beta, and they do not hide the downsides. The fine-tuned model skips words more often, and on ordinary speech it reproduces casing and punctuation worse. In addition, the test set was labeled by the same automatic pipeline as the training set, with selective manual checks. Part of the gain may come from matching labeling style rather than from real recognition improvement.
What to do with this
If you are building voice input, call transcription or an assistant for developers, you can take a few rules from this story:
- before writing a replacement dictionary, collect a list of real errors: if one term is spelled five different ways, fix the model;
- fine-tuning on tens of hours of live speech can double term accuracy without changing the architecture and without raising hardware load;
- keep compatibility with the weight format so that swapping and rolling back the model takes minutes;
- label the test set independently of the training set, otherwise you are measuring agreement with your own pipeline;
- watch for side effects: better term accuracy may cost you skipped words and damaged punctuation.
For narrow domain vocabulary such as IT terms, fine-tuning an existing model looks more sensible than trying to close the gap with dictionaries.
- Nov 8, 2026AI visit notes fill 50 fields in 35 seconds in Moscow, while private clinics remain silent
- Nov 21, 2026Whisper Handles 454 IT Terms Better Than GigaAM, Even Though GigaAM Makes Fewer Word Errors
- Oct 10, 2026Developer of GigaAM-Powered Offline Dictation App Brings Pisar to Google Play, Asks for Testers
- Oct 27, 2026SberDevices Releases Open-Source GigaAM-v3: Russian Speech Recognition at Least 10% Better
Key Questions Answered
What exactly was breaking?
The base model understood ordinary Russian speech well, but it failed on three groups of words: English names (Docker, Kubernetes), established loanwords such as "dashboard" in its Russian spelling, and the abbreviation for "AI" (written "II" in Russian). Sometimes whole words disappeared. The model transcribed the Russian sentence "My AI rarely makes mistakes" as "My rarely makes mistakes", dropping the abbreviation. A simple replacement dictionary only helps partly.
How it was fine-tuned and what came out?
GigaAM is a family of speech recognition models built on the Conformer architecture. The team uses the variant with an RNNT head, which outputs text with casing and punctuation directly. On the server, the encoder runs in FP16 and the RNNT head in FP32. Silero VAD cuts long recordings at pauses, and short segments are recognized in batches. Decoding is greedy, with no external language model, and the hardware is an RTX A2000 GPU with 12 GB.
What are the key details on Lesson from the first failed attempt?
The team first took the CTC variant of GigaAM and assembled 4.4 hours of data: about 1100 synthetic phrases with IT terms and roughly as many recordings from the Russian and English parts of the FLEURS dataset. They froze the bottom 12 encoder layers and trained the top four layers and the CTC head. On a test bench of 12 synthetic phrases, the share of correct terms jumped from 6% to 86%. But when compared with the production RNNT model on 150 live FLEURS recordings, the picture looked different.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.