GigaChat Audio detects emotions in long files

GigaChat Audio, an updated large language model, processes audio files without converting speech to text first. It detects positive or negative emotions from intonation and navigates recordings up to three hours long. The model achieved 80% accuracy in emotion recognition and a 70% win rate on the Arena-Hard-Audio benchmark. A lightweight open-source version, GigaChat3.1-Audio-10B, is available on GitVerse and Hugging Face.