Reproduce “Synaspot: A Lightweight, Streaming Multi-modal Framework for Keyword Spotting with Audio-Text Synergy” 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
Kewei Li, Yinan Zhong, Xiaotao Liang, Tianchi Dai, Shaofei Xue
SYNASPOT presents a lightweight, streaming multi-modal framework for open-vocabulary keyword spotting, leveraging audio-text synergy and a speaker-agnostic design for robust detection of varying keyword lengths in continuous speech. The framework, with 0.9M parameters, achieved a 5.77% EER on LibriPhrase ($LP_E$) and demonstrated competitive performance compared to larger baseline models in streaming scenarios.
Key insights
- Multimodal integration of audio, text, and combined audio-text features through contrastive learning can synergistically enhance KWS expressiveness and accuracy.
- Explicitly reducing speaker-specific information in audio embeddings via domain adaptation (gradient reversal) significantly improves model generalization across diverse speakers.
- A lightweight, streaming decoding mechanism relying solely on encoder representations and mathematical computations can effectively handle continuous speech and dynamically adapt to varying keyword lengths.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.