Sber updates GigaChat with audio model for voice tasks

Sber has updated its GigaChat AI assistant with a new audio model and released a lighter open-source version for developers.

The updated model processes audio files and voice messages directly, rather than converting speech to text first. It was trained to interpret intonation as well as spoken content, allowing it to respond to the tone of a user's voice.

That means the assistant can detect whether a speaker sounds positive or negative and adjust its replies accordingly. It can also retain facts disclosed in voice conversations for use in later sessions. Users can review, edit or disable that memory function in profile settings.

The model also supports longer recordings. According to Sber, it can handle audio files of up to three hours, navigate to points where a topic was discussed, summarise sections, generate full summaries with timestamps and distinguish between different speakers.

These functions could support meeting reviews, call analysis, note-taking and navigation through large volumes of recorded speech. The open-source version, called GigaChat3.1-Audio-10B, supports Russian, English and other languages.

Performance claims

Sber compared the model with rival systems on audio understanding and emotion recognition. It said GigaChat Audio recorded a 70% win rate on the Arena-Hard-Audio benchmark, compared to 77.5% for Gemini 3 Flash preview and 62% for Gemini 2.5 Pro.

On emotion recognition, Sber said its model achieved 80% accuracy, compared with 70% for Qwen3-Omni-30B and 62% for Kimi-Audio. Sber said the results show its audio assistant performs at a level close to leading international systems.

The developer release is aimed at users, including transcription, pronunciation training, voice-over assessment, context-aware interpreting and summarising long recordings. The code is available for free through developer platforms, including GitVerse and Hugging Face.

Multilingual push

Alongside the GigaChat release, Sber has made available GigaAM Multilingual, a family of automatic speech recognition models. The software supports Russian, English, Kyrgyz, Kazakh and Uzbek, and comes in a compact version for standard processors as well as a larger model with stronger recognition quality.

Sber said GigaAM makes between 1.5 and two times fewer errors than its nearest competitors in benchmark testing for Russian speech recognition. It said the model could be used in call centres, voice assistants, meeting transcripts, interviews, podcasts, voice input for applications, and automatic subtitle generation.

Sber added that both model families were pre-trained across multiple languages, which should make them easier to adapt to additional languages, including those used across the CIS. It said this fine-tuning would require only a few dozen hours of annotated audio.

Anton Frolov, Senior Vice President and Head of GenAI Development at Sberbank, outlined the company's view of the market for voice-based systems.

"Voice is the most natural way to interact with technology, but also the most demanding one, as any detection error or a misread emotion destroys your trust in the assistant immediately. That is why we believe that audio is where the AI assistant market has room for growth. By giving open access to these models, we are handing developers and researchers a powerful tool. The range of applications for voice technology is very wide, ranging from simultaneous voice translation to services for people with speech disorders, while multilingualism and the ability to easily learn other languages unlock international opportunities for these models," Frolov said.