A Posterior Approach for Microphone Array Based Speech Recognition

Automatic speech recognition (ASR) is difficult in environments such as multiparty meetings because of adverse acoustic conditions: background noise, reverberation and cross-talk. Microphone arrays can increase ASR accuracy dramatically in such situations. However, most existing beamforming techniques use time-domain signal processing theory and are based on a geometric analysis of the relationship between sources and microphones. This limits their application, and leads to performance degradation when the geometric properties are unavailable, or heterogeneous channels are used. We present a new posterior-based approach for microphone array speech recognition. Instead of enhancing speech signals, we enhance posterior phone probabilities which are used in a tandem ANN-HMM system. Significant improvements were achieved over a single channel baseline. Combining beamforming and our method is significantly better than beamforming alone, especially in a moving speakers scenario.