"Sber" has released an updated GigaChat Audio model for processing voice messages and audio files. It works directly with sound, without prior speech-to-text conversion, and determines whether the user is positive or negative based on intonation, voice, and pronunciation.

The neural network can analyze recordings up to three hours long and distinguish speakers within them. It can be asked to find the moment when a specific topic was discussed, retell a separate segment, or prepare a summary of the entire conversation with links to timestamps.

The AI assistant has also learned to remember facts from voice dialogues and take them into account in subsequent interactions. For example, when planning a travel route, "GigaChat" can rely on previously expressed interests. Saved information can be viewed, edited, or completely disabled in the settings.

According to Sber's internal tests, emotion recognition accuracy reaches 80%. In Arena Hard Audio, the model achieved 75% wins – almost on par with Gemini-3-Flash-preview at 77.5% and higher than Gemini-2.5-Pro at 62%. GigaChat Audio is already available within "GigaChat" and has been released for developers.