Reproduce “Query-by-example on-device keyword spotting” 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
Byeonggeun Kim, Mingu Lee, Jinkyu Lee, Yeonseok Kim, Kyuwoong Hwang
Qualcomm AI Research presents an on-device query-by-example keyword spotting system leveraging a small-footprint ASR model and Finite-State Transducers for phonetic matching. The system's key innovation is a method for robust on-device threshold prediction that utilizes query-specific negative examples generated directly from user-enrolled keywords, achieving consistent false alarm rates and competitive performance.
Key insights
- FST-based phonetic matching, when constrained by phonetic hypotheses, consistently and significantly outperforms Subsequence Dynamic Time Warping (S-DTW) for query-by-example keyword spotting.
- Generating 'query-specific negatives' by shuffling parts of user enrollment utterances provides an effective and robust method for on-device threshold prediction, addressing the challenge of limited negative examples.
- A small-footprint Acoustic Model combined with FSTs allows user-specific keyword spotting to achieve performance comparable to state-of-the-art predefined KWS systems while operating on-device.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.