🤖 Sber has introduced new speech recognition models — GigaAM Multilingual and GigaChat Audio
Sber has opened access to the GigaAM Multilingual and GigaChat Audio model families. GigaAM Multilingual, based on Conformer, was trained on 2 million hours of audio in 70+ languages, including Kazakh, Kyrgyz, and Uzbek, outperforming Whisper Large v3 in these regions. GigaChat Audio (10B parameters, MoE) combines speech recognition with GigaChat 3.1 capabilities to process recordings up to 2 hours long.
🌍 The release of high-quality open-source models optimized for CIS languages reduces dependence on Western solutions and enables the construction of efficient systems for regional markets.
👤 Professional audio analysis tools can now be used for free by downloading the model weights from Hugging Face.
Source 1: https://huggingface.co/ai-sage/GigaAM-Multilingual Source 2: https://huggingface.co/ai-sage/GigaChat3.1-Audio-10B-A1.8B