Speech estimation from body-conducted speech with differential acceleration

Speech-recognition rates decrease in noisy environments. The body-conducted speech, conducted in solids such as body and skins, has a noise-robust characteristics and can be served for recognition systems even in 98 dBSPL (-20 dBSNR) noise environments. However, the body-conduction can not capture high frequency sounds. Conventional methods to improve sound quality of body-conducted speeches need both speeches themselves and body-conducted speeches. In this paper, a new body-conducted speech retrieval technique in sound quality without a speech signal itself is proposed. First, high-frequency components in the body-conducted speech were emphasized using differential acceleration. Second, a conventional noise reduction method was adopted to make a clear body-conducted speech from a retrieval speech which contains constant noise. The recognition experiments using the proposed method showed that it improved recognition rate in all speakers.