GigaChat Audio: Time-aware Large Audio Language Model

Researchers at SaluteDevices present GigaChat Audio, an open-weight time-aware Audio LLM built on a 10B-A1.8B mixture-of-experts language model that processes recordings up to 120 minutes and generates answers, descriptions, and summaries anchored to explicit timestamps. Its central technique interleaves continuous audio tokens with periodic temporal anchors, using a cascaded synthetic-supervision pipeline: WhisperX aligns YODAS2 transcripts, GPT-OSS-120B generates timestamped questions and answers from roughly 10-minute slices, and a global verifier filters inconsistencies. On 20–40-minute temporal grounding, the model reaches 53.8 mIoU with anchors every 60 seconds, while removing anchors drops performance to 14.2 mIoU; increasing anchor frequency to every 7 seconds raises accuracy to 65.2 mIoU. At 60-second spacing, midpoint timing error is 3 seconds, showing that sparse anchors can support fine-grained localization through interpolation. The training corpus retains 14k hours after filtering from multilingual YODAS2 audio and mixes recordings across duration ranges, because short-only training fails to extrapolate to long recordings and long-only training harms short-audio performance. Compared with systems such as OpenAI’s GPT-OSS-120B used for data generation, Qwen3-Omni, and Google’s Gemini 3 Flash, the work emphasizes verifiable long-form temporal grounding rather than general audio question answering, and releases model weights plus a 10k+ hour temporal dataset.