Sixth International Conference on Spoken Language Processing
Cepstral mean subtraction (CMS), which is a simple long-term bias removal, is used to compensate for transmission and linear fixed channel effects. In order to process the non-linear channel, a two-level CMS was proposed where separate channel compensation is performed for segments that are classified as speech and for segments classified as background. In this paper, methods for extending the two-level CMS to real-time implementation is proposed using a finite number of look-a-head frame delay, which further reduces computation and memory requirements of the compensation process. The on-line bias compensation shows similar characteristic curve as that of batch-mode and has the effect of greatly reducing the sensitivity of the recognizer to transmission noise variability.
Bibliographic reference. Chengalvarayan, Rathinavelu (2000): "Look-ahead sequential feature vector normalization for noisy speech recognition", In ICSLP-2000, vol.4, 524-527.