Combining the SNR Spectrum with a Cochlear Model

The SNR spectrum was previously introduced as a natural consequence of using cepstral normalisa- tion in speech recognition; it is closely related to the articulation index of Fletcher. Motivated initially by a theoretical difficulty in frequency warping, the SNR spectrum is combined with a cochlear model, yielding a theoretically sound yet simple noise robust feature. Two time-domain cochlear model imple- mentations, representing different recursive approximations, are compared. The combination of SNR spectrum, cochlear model, perceptual linear prediction and cepstral normalisation leads to an intuitive, efficient and effective feature that is mostly physiologically plausible.