Improvement of robustness using selective sound segregation for automatic speech recognition systems in noisy environments (Speech) -- (国際ワークショップ"Asian workshop on speech science and technology")

概要

論文の詳細を見る
This paper proposes the concept of our novel robust speech recognition method based on the selective sound segregation model, and demonstrates that the proposed method can play an effective role to improve robustness of automatic speech recognition (ASR) systems in various noisy environments. Almost all ASR systems for noise environments attempt to transform an input sound into a clean speech or reference patterns into ones adapted for noises using a noise model, and calculate similarity between an input sound and reference patterns. In our proposed method, the possibility of existence of a target speech in an input sound is employed as a measure of recognition. The possibility of existence of a target speech is calculated by validity of the selective sound segregation model without any noise model. An ASR system based on our proposed method was implemented. To evaluate our proposed ASR system, Japanese digits recognitions in various noisy environments were carried out using traditional ASR systems and the proposed one. Results showed that the proposed method is more robust than other in experimental conditions in SNR = 0 dB. These indicate the proposed method can play an effective role to improve robustness of the ASR systems.
社団法人電子情報通信学会の論文
2008-03-13