Technical writing
Notes, tasks, documentation, and emails where Docker and PostgreSQL should remain recognizable names.
Speech recognition adaptation · 0.02
Russian speech.
English
technical terms.
Kamertune adapts GigaAM-v3 to Russian speech with English technical vocabulary. The model is already used in Glas.
We measure terminology recognition alongside word errors, capitalization, and punctuation.
Technical speechRU + EN
Illustrative target text
Обнови конфиг Nginx и перезапусти Docker.
Illustrative spelling, not a new model test result.
92.2%IT-term recall
5.12%WER — word error rate
525held-out dictations in the test
Evaluation: Nov 8, 2026 · Internal held-out set · Term recall is not overall transcription accuracy
01 / Results
Progress you can measure.
The original model and two Kamertune versions, evaluated on the same set of short dictations.
| Metric | GigaAM-v3 original | Kamertune 0.01 | Kamertune 0.02 |
|---|---|---|---|
| WER lower is better | 8.77% | 5.80% | 5.12% |
| WER incl. case and punctuation lower is better | 10.51% | 8.88% | 8.25% |
| IT-term recall higher is better | 44.5% | 86.7% | 92.2% |
WER counts word substitutions, deletions, and insertions. IT-term recall is the share of target terms recognized correctly. These are internal test results, not a guarantee for every recording.
02 / Use cases
Talk about your work.
Keep the technical vocabulary.
Russian speech often includes English names. Especially when the subject is software, products, and technology.
Instructions for AI tools
Detailed requests and follow-ups that mix Russian sentences with FastAPI, pull request, and library names.
Speech in your product
For teams building products for Russian-speaking users with technical vocabulary. API access information will be added to the Glas website.
03 / Method
Adapt the model.
Evaluate the whole transcript.
Kamertune 0.02 continues fine-tuning GigaAM-v3 RNNT on real speech. Technical vocabulary is evaluated alongside punctuation and the rest of the text.
Real speech
Dictations and other speech recordings with technical vocabulary, rather than synthetic phrases alone.
Consistent spelling
Names and abbreviations in Latin script. Established Russian loanwords keep their familiar spelling.
Held-out evaluation
Separate speakers and recordings. Candidates are compared with the previous and original models.
Multiple metrics
WER, case and punctuation, and term recall. A single headline metric cannot tell the whole story.
71.7 hours of fine-tuning for 0.0225,327 training recordingsUsed in Glas since Nov 8, 2026
04 / Limitations
Adapted to a specific task.
Errors still happen.
Focused on Russian technical speech
Kamertune targets Russian speech containing English technical terms. This does not imply equal performance on unrestricted English speech, every accent, or noisy recordings.
Not every test improves
In the same evaluation, FLEURS ru WER (150 recordings) changed from 5.01% to 5.11%; clean synthetic speech (36 recordings) changed from 2.92% to 3.51%, for 0.01 → 0.02. These sets are measured separately from dictations.
Results come from the internal Kamertune acceptance evaluation on November 8, 2026. Metrics are from a single run and are not mixed with other evaluations. For the base architecture, see the official GigaAM project.
Next step
From measured results
to your own workflow.
Dictate with Glas
Try recognition on your own speech and see how much editing is needed in your day-to-day work.
Try GlasFor your product
Interested in speech recognition through an API? Information about access will be published on the Glas website.
Visit GlasMore detail
Questions, answered
How are Kamertune, GigaAM, and Glas related?
GigaAM-v3 is the base speech recognition model. Kamertune adapts it to Russian speech with technical vocabulary. Glas is the application that uses the resulting model.
Does 92.2% mean overall transcription accuracy?
No. It is target IT-term recall on a specific held-out set of 525 dictations. Errors across the whole transcript are measured separately using WER.
Can I download the Kamertune code and weights?
This page presents results and the evaluation approach. Kamertune code, data, and weights are not published for download. The original GigaAM has its own public repository.
How can I use the API?
Information about API access will be published on the Glas website. Please check Glas for availability, terms, and integration details.
What do the examples above show?
They illustrate target spelling for English terms in a Russian sentence. They are not model responses from a new test. Measured evaluation results are shown separately.