Improvement of body-conducted speech recognition using model estimation

One problem with speech recognition is a low performance in noisy environments because it is easily influenced by aerial in air. Although the sound quality of body-conducted speech (BCS) and regular speech are different, with BCS recognition, it is possible to recognize an utterance in noisy environments with a rating of 98 dB sound pressure level (SPL) in our previous study. In this study, we investigate how to improve BCS recognition performance using model re-estimation methods of ML and MAP. An acoustic model uses parameters such as mean vector, covariance matrix, weight, and transition probability. Recognition performance is improved by model re-estimation of speech and BCS using maximum likelihood and maximum a posteriori methods, respectively. We confirmed that improvements in recognition performance are achieved for practical through the re-estimation of the covariance matrix and mean vector.