Improvement of throat speech quality for speech recognition using Bayes' theorem

In general, a speech signal can be measured by a microphone, such as a throat microphone. However, the speech signal measured by a microphone often contains surrounding noise. On the other hand, although a throat microphone is effective for surrounding noise, the speech signal it measures includes body-conducted internal noise. In this study, we propose an improvement method for the sound quality of the speech signal measured by a throat microphone to achieve speech recognition well. The relationship between the original speech signal and the speech measured by the throat microphone is not clear. Therefore, we consider the relationship as a multiplicative and additive model of the original speech signal and noise components with unknown parameters. An algorithm is proposed to simultaneously estimate the original speech signal and the unknown parameters using Bayes’ theorem based on the speech signal measured by the throat microphone. Finally, a speech recognition experiment is conducted to confirm the effectiveness of the proposed algorithm.