Saved in:
Bibliographic Details
Main Authors: Ai, Zhiqi, Cheng, Han, Wang, Yuxin, Mu, Shiyi, Xu, Shugong, Zhou, Yongjin
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.10740
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908589055016960
author Ai, Zhiqi
Cheng, Han
Wang, Yuxin
Mu, Shiyi
Xu, Shugong
Zhou, Yongjin
author_facet Ai, Zhiqi
Cheng, Han
Wang, Yuxin
Mu, Shiyi
Xu, Shugong
Zhou, Yongjin
contents In this paper, we propose DS-KWS, a two-stage framework for robust user-defined keyword spotting. It combines a CTC-based method with a streaming phoneme search module to locate candidate segments, followed by a QbyT-based method with a phoneme matcher module for verification at both the phoneme and utterance levels. To further improve performance, we introduce a dual data scaling strategy: (1) expanding the ASR corpus from 460 to 1,460 hours to strengthen the acoustic model; and (2) leveraging over 155k anchor classes to train the phoneme matcher, significantly enhancing the distinction of confusable words. Experiments on LibriPhrase show that DS-KWS significantly outperforms existing methods, achieving 6.13\% EER and 97.85\% AUC on the Hard subset. On Hey-Snips, it achieves zero-shot performance comparable to full-shot trained models, reaching 99.13\% recall at one false alarm per hour.
format Preprint
id arxiv_https___arxiv_org_abs_2510_10740
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Dual Data Scaling for Robust Two-Stage User-Defined Keyword Spotting
Ai, Zhiqi
Cheng, Han
Wang, Yuxin
Mu, Shiyi
Xu, Shugong
Zhou, Yongjin
Sound
In this paper, we propose DS-KWS, a two-stage framework for robust user-defined keyword spotting. It combines a CTC-based method with a streaming phoneme search module to locate candidate segments, followed by a QbyT-based method with a phoneme matcher module for verification at both the phoneme and utterance levels. To further improve performance, we introduce a dual data scaling strategy: (1) expanding the ASR corpus from 460 to 1,460 hours to strengthen the acoustic model; and (2) leveraging over 155k anchor classes to train the phoneme matcher, significantly enhancing the distinction of confusable words. Experiments on LibriPhrase show that DS-KWS significantly outperforms existing methods, achieving 6.13\% EER and 97.85\% AUC on the Hard subset. On Hey-Snips, it achieves zero-shot performance comparable to full-shot trained models, reaching 99.13\% recall at one false alarm per hour.
title Dual Data Scaling for Robust Two-Stage User-Defined Keyword Spotting
topic Sound
url https://arxiv.org/abs/2510.10740