Unvoiced Speech Recognition Based on One-Channel Facial Myoelectric Signal

Abstract

Six Chinese vowels /a/, /o/, /e/, /i/, /u/, and / / are recognized based on the one-channel detected facial myoelectric signal (MES). Zygomaticus major and anterior belly of the digastric are carefully selected as the electrodes sites of MES detected. Over-sampling technology and four-layer wavelet decomposition are used to reduce noise in MES records. Digital down converter down converts the original sampling rate to Nyquist frequency. By anterior methods, clean MES is gotten with no signal distortion. For MES is not voice signal, MFCC is not selected as the feature set. According to MES characteristic, ten-order AR model is set up. The coefficients of AR model, cepstral coefficients, and amplitude of MES are chosen to form the original feature set for recognition. Principal components analysis (PCA) reduces the dimension of original feature set before the proposed BP networks. Combining two BP network classifier, an efficient classifier is proposed. The result of experiment shows that Chinese vowel / /, /and /i/ have good classification rate (more than 90%) based on one-channel facial myoelectric signal. Index Terms -unvoice speech, feature extraction, BP network, endpoint detection, wavelet decomposition.