SberDevices Releases Open-Source GigaAM-v3: Russian Speech Recognition at Least 10% Better
SberDevices has released GigaAM-v3, an open-source model that turns Russian speech into text with punctuation and numbers already in place. Its new HuBERT-CTC pretraining uses a ready-made speech recognition model as a teacher and delivers a quality gain of at least 10% over common methods. The labeled recordings have doubled to 4,000 hours, including call-center audio. For Russian voice services, it's a ready-made foundation.
AI-processed from SberDevices; edited by Hamidun News
If you're building a voice assistant, transcribing calls, or generating subtitles for Russian-language video, you don't just need a model that hears words. You need readable text: with commas, question marks, and digits instead of "twenty-five percent." SberDevices has presented GigaAM-v3, the new version of its open family of acoustic models for Russian, and put the emphasis squarely there. The authors' main claim: the new HuBERT-CTC pretraining method delivers a quality gain of at least 10% even compared with methods common in the industry.
How the Encoder Is Trained: HuBERT-CTC
Modern automatic speech recognition (ASR) models are first pretrained on huge amounts of unlabeled audio. The idea resembles BERT for text: part of the audio is masked, and the model guesses what was there. In classic HuBERT, the "correct answers" for this guessing game are produced as follows: acoustic features are taken, grouped with k-means, and turned into a set of discrete tokens, a kind of "sound alphabet."
Sber changed the starting point. Instead of a generic encoder, they took an encoder that had already been fine-tuned for speech recognition with a CTC objective. Its internal representations capture phonetics and meaning better, so the "sound alphabet" ends up closer to the real task. From there, everything works as in HuBERT: fragments are masked and the model learns to reconstruct the tokens. Put simply, the knowledge of a ready-made ASR model flows into pretraining, and a very large volume of unlabeled audio is still used.
What the experiments showed. In a test with a frozen encoder, recognition quality improves from the lower layers toward the upper ones. For the competing BEST-RQ method, the minimum word error rate (WER) is often reached in the middle layers. The final WER after fine-tuning is lower than that of plain HuBERT, even when the full labeled dataset is used. When labeled data is scarce, at 10%, 1%, or 0.1% of the full set, the gain is more pronounced.
Data: More Hours and More Situations
GigaAM-v2 was pretrained on roughly 50,000 hours of Russian speech, and the volume was increased for v3. Recordings are cleaned in a simple way: a voice activity detector (VAD) cuts long files into minute-long fragments and discards those with too much silence. The team tried more sophisticated filters, but they barely affected quality and required significant compute, so they kept VAD.
The labeled data has roughly doubled, to 4,000 hours, and it is more diverse:
- complex texts on which other models made mistakes: 950 hours;
- spontaneous conversational speech: 450 hours;
- call centers and telephone recordings: 350 hours;
- speech against background music: 110 hours;
- synthesized speech: 100 hours;
- speech with special characteristics: 50 hours.
On open datasets (Golos, Common Voice, Russian LibriSpeech), quality stayed at the v2 level. The authors explain this by a marked increase in the share of difficult internal domains. That's why they separately report results on call-center recordings and on speech with special characteristics. For practitioners, this is an honest signal: on "clean" benchmarks, v3 does not outperform v2, and the gain has to be sought in difficult conditions.
Readable Text Instead of Service Transcripts
Standard datasets store transcriptions in a service format: without punctuation, with numbers and symbols such as % and $ written out as words. That's convenient for training but not for reading. To obtain a corpus with normal text, the team used GigaChat Audio. The model was given the audio, the original transcript, detailed instructions, and 10 fixed examples. The audio is needed so the model can use intonation to recognize a question and handle pauses correctly.
Generative models hallucinate, so the output was filtered:
- they checked that the text length barely changed;
- they removed characters outside the defined alphabet, such as emoji;
- they stripped Markdown and inserts like "(silence)";
- they verified that the % sign appeared only where the original said "percent," and $ only where it said "dollar";
- they monitored the divergence on CER and WER metrics.
An alternative was forced decoding: punctuation is added directly during decoding, and the reference text is left unchanged, so WER cannot get worse. The drawback of this method is reduced flexibility in normalization, and the resulting punctuation is worse than with LLM-based labeling. The approaches were compared on internal data, where transcription and punctuation were labeled by human assessors.
What to Do With This
GigaAM-v3 is an open model, so you can run it yourself and avoid sending sensitive calls to external APIs. But the lesson goes beyond the model itself. If you have raw transcripts, an LLM with access to the audio and strict filters can turn them into training data with punctuation. And when choosing an ASR system, look beyond WER on open datasets to your own domains: noisy calls, music, and non-standard speech. For evaluation across several domains, the authors themselves added LLM-as-a-judge, scoring by another language model, on top of classic WER, because a single share of erroneous words poorly reflects the readability of text with punctuation.
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
- Nov 21, 2026Waymo pulls robotaxis off highways after new entries into work zones
- Nov 21, 2026Google Opens Pre-Orders for Gemini Smart Glasses from $395, Shipping December 10
- Nov 21, 2026Space Data Centers Are Still Unprofitable: Launch Costs Run 3–19 Times Over the Break-Even Point
Key Questions Answered
How the Encoder Is Trained: HuBERT-CTC?
Modern automatic speech recognition (ASR) models are first pretrained on huge amounts of unlabeled audio. The idea resembles BERT for text: part of the audio is masked, and the model guesses what was there. In classic HuBERT, the "correct answers" for this guessing game are produced as follows: acoustic features are taken, grouped with k-means, and turned into a set of discrete tokens, a kind of "sound alphabet." Sber changed the starting point.
What are the key details on Data: More Hours and More Situations?
GigaAM-v2 was pretrained on roughly 50,000 hours of Russian speech, and the volume was increased for v3. Recordings are cleaned in a simple way: a voice activity detector (VAD) cuts long files into minute-long fragments and discards those with too much silence. The team tried more sophisticated filters, but they barely affected quality and required significant compute, so they kept VAD.
What are the key details on Readable Text Instead of Service Transcripts?
Standard datasets store transcriptions in a service format: without punctuation, with numbers and symbols such as % and $ written out as words. That's convenient for training but not for reading. To obtain a corpus with normal text, the team used GigaChat Audio. The model was given the audio, the original transcript, detailed instructions, and 10 fixed examples. The audio is needed so the model can use intonation to recognize a question and handle pauses correctly.
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.