1-1hit |
Yoshinori MORITA Tetsuo FUNADA Hideyuki NOMURA
The bandwidth occupied by individual telecommunication devices in the field of mobile radio communication must be narrow in order to effectively exploit the limited frequency band. Therefore, it is necessary to implement low-bit-rate speech coding that is robust against background noise. We examine vector quantization using a neural network (NNVQ) as a robust LSP encoder. In this paper, we compare four types of binary patterns of a hidden layer, and clarify the dependency of quantization distortion on the bit pattern. By delayed decision (selection of low-distortion codes in decoding, i.e., EbD method) the spectral distortion (SD) can be decreased by 0.8 dB (20%). For noisy speech, the performance of the EbD method is better than that of the conventional VQ codebook mapping method. In addition, the SD can be decreased by 2.3 dB (40%) by using a method in which the neural networks for encoding and decoding are combined and re-trained. Finally, we examine the SD for speech having different signal-to-noise ratios (SNRs) from that used in training. The experimental results show that training using SNR between 30 and 40 dB is appropriate.