Noise masking in a transform domain

A.P. Varga and K.M. Ponting (1989) have demonstrated that the noise masking algorithm can provide effective noise compensation, leading to robust speech recognition with almost undegraded performance down to signal-to-noise ratios as low as 3 dB in stationary pink noise. Noise compensation techniques have been limited to the log energy spectral domain, whereas recognition is more normally carried out in a transform domain. The authors describe the adaptation of the noise masking algorithm for use in a cepstral transform domain, providing performance which is not significantly different from masking in the log energy domain. On the task examined here the technique compares favourably with the more computationally demanding hidden Markov model decomposition approach, with the performances diverging significantly only for signal-to-noise ratios worse than 3 dB. It can be seen that the use of noise masking can provide a good degree of noise robustness in both the filter bank domain and the mel scaled frequency cepstral domain.<>