Sber has introduced the GigaAM Multilingual and GigaChat Audio model families, providing developers with open access to powerful speech recognition and analysis tools optimized for CIS languages.
What Happened
Sber has released GigaAM Multilingual, a model built on the Conformer architecture and trained on 2 million hours of audio across more than 70 languages, as well as the multimodal GigaChat Audio model (10B parameters, MoE architecture). GigaAM Multilingual outperforms Whisper Large v3 on languages such as Kazakh, Kyrgyz, and Uzbek. GigaChat Audio integrates speech recognition capabilities with the GigaChat 3.1 language model, allowing for the processing of recordings up to 2 hours long and providing accurate event localization.
Context
GigaAM Multilingual utilizes a compact architecture (240M parameters) for high efficiency on regional languages, while GigaChat Audio demonstrates high event localization accuracy (IoU 48.3 on recordings lasting 20–60 minutes), making it applicable for deep transcription and audio content searching.
Why It Matters for the Industry
The release of high-quality open-source models reduces the industry's dependence on Western solutions like Whisper and allows local companies to build effective speech processing systems for CIS markets. The availability of weights on Hugging Face and code on GitHub significantly lowers the technological barrier and development costs for AI-audio startups.
Why It Matters for Users
Developers and engineers can use professional speech recognition and audio analysis tools for free, including searching for specific moments in long recordings, by integrating them into their local pipelines without the need to use foreign APIs.
Sources
- ai-sage/GigaAM-Multilingual · Hugging Face
- ai-sage/GigaChat3.1-Audio-10B-A1.8B · Hugging Face
- salute-developers/GigaAM · GitHub
Author
Look at AI, Editorial Team