Acoustical Pre-Processing for Robust Speech Recognition.

摘要

In this paper we describe our initial efforts to make SPHINX, the CMU continuous speech recognition system, environmentally robust. Our work has two major goals: to enable SPHINX to adapt to changes in microphone and acoustical environment, and to improve the performance of SPHINX when it is trained and tested using a desk-top microphone. This talk will describe some of our work in acoustical pre-processing techniques, specifically spectral normalization and spectral subtraction performed using an efficient pair of algorithms that operate primarily in the cepstral domain. The effects of these signal processing algorithms on the recognition accuracy of the Sphinx speech recognition system was compared using speech simultaneously recorded from two types of microphones: the standard close-talking Sennheiser HMD224 microphone and the desk-top Crown PZM6fs microphone. A naturally- elicited alphanumeric speech database was used. In initial results using the stereo alphanumeric database, we found that both the spectral subtraction and spectral normalization algorithms were able to provide very substantial improvements in recognition accuracy when the system was trained on the close-talking microphone and tested on the desk-top microphone, or vice versa. Improving the recognition accuracy of the system when trained and tested on the desk-top microphone remains a difficult problem requiring more sophisticated noise suppression techniques.