Using noise reduction and spectral emphasis techniques to improve ASR performance in noisy conditions

The proposed algorithm uses noise reduction and spectral emphasis techniques to get more robust features when the input speech is distorted by various noises or channel distortions. The AURORA 2.0 Database, together with the HTK speech recognition toolkit, is used to evaluate the performance of speech recognition with the proposed algorithm. Both noise reduction and spectral emphasis subroutines are implemented in the baseline front-end processing program with a low computational load and for real-time operation. With the integration of a voice activity detector, it is shown that the proposed algorithm improves the recognition results by 46.54% over the reference front-end algorithm in clean-condition training. It uses existing HMMs trained by clean speech and it works well on all testing cases.