A Frequency-Weighted Itakura Spectral Distortion Measure and Its Application to Speech Recognition in Noise

The Itakura LPC spectral distortion measure, or the log likelihood ratio distortion measure, has been applied successfully to the recognition of noise-free or low-noise speech signals. However, the performance of a recognizer based on this distortion measure deteriorates when speech signals are corrupted by noise. In certain applications, where training and testing can be done under the same noise conditions, the degradation of performance is usually small. For other applications, where it is not feasible to train and to test the recognizer under similar noise conditions, the performance is often unacceptable. In such situations, a more noise-resistant distortion measure is required. Since, in general, the speech power, is nonuniformly distributed along the frequency axis, we consider a weighted spectral distortion measure which weights the high SNR regions in frequency more than the low SNR regions. The weighting is also adjusted adaptively according to an SNR measurement.