GigaChat Audio processes 2-hour audio with precise timestamps
New time-aware LLM answers questions with explicit timestamps over 120 minutes.
GigaChat Audio tackles a core limitation of audio-conditioned LLMs: temporal grounding in long recordings. The model, developed by authors from the GigaChat team (likely Sber AI), interleaves periodic time markers with continuous audio tokens using large-scale synthetic supervision from a cascaded pipeline. This design lets it answer questions with explicit timestamps across inputs up to 120 minutes long, a dramatic improvement over prior models that struggle with even short clips.
The paper reports strong temporal-grounding accuracy on both short and long benchmarks, and the model supports time-anchored fragment descriptions and summaries. Extensive ablations examine how time representation, marker frequency, tokenization, and duration-mixture design affect accuracy and computational cost. Accepted to Interspeech 2026, the authors have released model weights and datasets to drive further research in time-aware audio understanding.
- Handles audio inputs up to 120 minutes long with explicit timestamp answers
- Achieves strong temporal-grounding accuracy on short and long benchmarks
- Open-sourced model and datasets, accepted to Interspeech 2026
Why It Matters
Enables precise search, summarization, and analysis of hour-long audio recordings with exact temporal context.