Reproduce “Listening Between the Frames: Bridging Temporal Gaps in Large Audio-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hualei Wang, Yiming Li, Shuo Ma, Hong Liu, Xiangdong Wang
TimeAudio uses temporal markers and absolute-time encoding to attach positions to audio features, while segment-level token merging reduces redundant tokens. It is trained on time-sensitive audio tasks using a 7B architecture. On 997 AudioGrounding queries, it reaches 57.8 mIoU versus 51.9 for FTAR-tuned 7B SALMONN, a reported 11.4% gain, but METEOR is lower than Qwen2-Audio on dense captioning.
Key insights
- Converting continuous timestamps into discrete, vocabulary-embedded temporal tokens (Temporal Markers) significantly improves the LLM's ability to generate precise time information.
- Explicitly injecting absolute time embeddings into audio features effectively grounds acoustic content to the timeline, preserving temporal order and specific event positions.
- An attentive, cluster-based token merging strategy can efficiently reduce the redundancy in long audio sequences, enabling end-to-end processing while retaining crucial semantic information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.