높은 소음 환경에서 강건한 음성 인식을 위한 ASR 시스템 개발

Abstract

Speech recognition is a task that takes speech as input and outputs text. Speech recognition is used in various places in the real world, such as AI speakers, voice memos, and more, and there is room for further development. For this purpose, various studies based on deep learning are being conducted. However, these studies are trained with data containing relatively quiet noises rather than real-world environments with strong noises including ambient sounds or noisy situations, resulting in poor recognition rates in noisy environments. We improved the performance of the Whisper model by using a pre-trained model that removes the noise through a process of compressive restoration, and showed that the method of learning by synthesizing noisy speech with denoised speech shows better performance than the model that learned only noisy speech, showing that the proposed method is effective for DNN-based speech models, and proposes a framework for creating noise robust speech recognition models.