A weight estimation method using LDA for multi-band speech recognition

This paper proposes a band-weight estimation method using Linear Discriminant Analysis (LDA) for multi-band automatic speech recognition (ASR). In our scheme, a spectral domain feature, SPEC, is modeled using a multi-stream HMM technique. This paper also proposes the use of Output Likelihood Normalization (OLN) in combination with the LDA-based weight-estimation method in order to adjust the relative weights of individual word (phoneme) models. Experiments were conducted using Japanese connected digit speech in various kinds of noise and SNR conditions. Experimental results show that the proposed LDA-based method is effective in all noise conditions. The results also confirm that the combination of OLN with the LDA-based method further increases noise robustness of the multi-band ASR. Furthermore, comparing the results of LDA applied to the SPEC and MFCC features respectively, it can be seen that greater performance gains are achieved with the former case than with the latter; this means that SPEC within a multi-band speech recognition framework can more effectively deal with the noise contamination than MFCC.