Evaluation and Modification of Cepstral Moment Normalization for Speech Recognition in Additibe Babble Ensemble

The statistical properties of a speech feature could differ under the influence of noisy environments. These effects are common in mismatched environments such as additive background noise and reverberant environments. Normalization strategies are employed in speech recognition systems to compensate for the effects of environmental mismatch. This work explores the utilization of cepstral moment normalization for speech recognition in additive noisy environment. The author has evaluated and modified the models used for cepstral moment normalization for improved convergence and performance. The cepstral moment normalization schemes were adopted for additive noisy speech recognition on TI-digit database. Further experimental works on cepstral moment normalization involving dynamic features are also presented. Both odd and even order cepstral moment normalization have shown significant contributions to speech recognition in low signal-to-noise ratio non-stationary noisy environment.