It is demanded that the occupied bandwidth of individual telecommunications devices is narrowed in the field of the mobile radio communication to exploit a limited frequency band effectively. For this purpose, low bit rate speech coding which has robustness against background noise is necessary. We examine vector quantization using a neural network (NNVQ) as robust LSP encoder. In this paper, we compare four kinds of hidden layer code patterns, and clarify the dependency of quantization distortion on the code pattern. LSP error can decrease by 0.007 (22) by choosing the code which yields low distortion in decoding (EbD method). To noisy speech, the EbD method shows that its performance is improved comparing with the conventional VQ method. LSP error can also decrease by 0.020 (43) by using such a method as the neural networks of encoding and decoding are combined and re-trained. Finally, we examine LSP error to the speech of different SNR from training. The experimental results show that training by the SNR of 30 to 40dB is appropriate.
展开▼
