COMPARISON OF NOISE ROBUST METHODS IN LARGE VOCABULARY SPEECH RECOGNITION

In this paper, a comparison of three fundamentally different noise robust approaches is carried out. The recognition performances of multicondition training, Data-driven Parallel Model Combination (DPMC), and cluster-based missing data reconstruction methods implemented in a large vocabulary continuous speech recognition system are evaluated with Finnish language speech data consisting of real recordings in noisy environments. All three methods improve the recognition accuracy substantially in poor signal-to-noise ratio (SNR) conditions when compared to a baseline system trained on clean speech. DPMC and missing data reconstruction systems give the best performance on high SNR conditions. On low SNR conditions, the performance of multicondition trained system is ranked the best, DPMC the second best and missing data reconstruction the third.