Target Speaker Extraction for Customizable Query-by-Example Keyword Spotting

To achieve a better user experience, it is desirable to have a customizable keyword spotting (KWS) system. Query-by-Example (QbE) is a promising way to achieve customization. In order to reduce the false alarms caused by interfering speech and ambient noise, we propose a speech enhancement frontend based on VoiceFilter for QbE based KWS system. VoiceFilter is a speaker extraction model, which extracts the voice of a target speaker from multi-speaker speech signals, with a reference signal from the target speaker. In this paper, we improve VoiceFilter substantially to better fit the KWS scenario, enhancing the voice of the target speaker, suppressing the voice of non-target speakers, and reducing ambient noise as well. To further reduce false rejections of the system with a VoiceFilter frontend, we apply exemplar augmentation to add reverberation to enrollment templates. Our proposed method leads to improved performance according to our experiments. Comparing with a DTW-based QbE system, our best system achieves a 39.0% relative reduction in false reject rate, at a false alarm rate of 0.5 times per hour.