A study of making clear body‐conducted speech using differential acceleration

Abstract

AbstractAn existing problem for speech recognition is that the recognition rate tends to decrease because of the noise. Body‐conducted speech offers a robust signal extraction method to extract speech signals from noisy environments, particularly because body‐conducted speech is a propagated sound and not easily influenced by noise. However, when body‐conducted speech is extracted with an accelerator, the typical frequency component of 2 kHz or more decreases in comparison with normal speech. Therefore, much research has been carried out investigating the techniques for improving the sound quality of body‐conducted speech. Particularly, conventional methods require normal speech to estimate clear speech. Here, we propose a technique that combines differential acceleration and noise reduction to estimate a clear signal using only body‐conducted speech. By comparing spectrogram differences with each method, we found that the Wiener filtering method was suitable for eliminating noise with differential acceleration. Thereafter, improvements in recognition rates were obtained using retrieval signals although the decoder has an acoustic model made by speech parameters. © 2011 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.