Data-Augumention Strategy For Small-Footprint Keyword Spotting

Due to the complex and variable environments of keyword spotting in real usage scenarios, there are multiple interference factors such as point-source noises, reverberation, background noises, dynamic changes in speaker's speed and volume. However, the current open source speech dataset lacks real data in such specific application scenarios. To solve this problem, we propose an efficient data modeling method in the real environment to process the data in the open source dataset to simulate the real use environment, so as to improve the robustness of the keyword spotting models. We evaluated our proposed data-augmentation strategy on the Mobvoihotwords dataset. Our results show that compared with the baseline system, the false rejection rate is reduced by 5.1% under the premise of a fixed false alarm per hour.