Dynamic feature variance adaptation for robust speech recognition with a speech enhancement pre-processor

摘要

It is well known that the performance of automatic speech recognition degrades severely in presence of noise or reverberation. Speech enhancement techniques may reduce such acoustic perturbations, but often do not interconnect well with speech recognizer. To cope with this problem, model adaptation is usually used to reduce the mismatch between the speech enhanced features and the acoustic model used by the recognizer. However, conventional model adaptation techniques assume static mismatch and may therefore not cope well with dynamic mismatch arising from noise or reverberation. There seems to be a lack of optimal ways to combine model adaptation and speech enhancement. In this paper we propose a novel adaptation scheme that may cope with dynamic mismatch. We introduce a parametric model for variance adaptation that includes static components, and dynamic components derived from a speech enhancement pre-process. The model parameters are optimized using adaptive training. An evaluation of the method with a speech dereverberation for pre-processing revealed that a 80% relative error rate reduction was possible compared with the recognition of dereverberated speech, and the final error rate was 5.4% which is close to that of clean speech (1.2%).