Evaluation of PNCC and extended spectral subtraction methods for robust speech recognition

This paper evaluates the robustness of different approaches for speech recognition with respect to signal-to-noise ratio (SNR), to signal level and to presence of non-speech data before and after utterances to be recognized. Three types of noise robust features are considered: Power Normalized Cepstral Coefficients (PNCC), Mel-Frequency Cepstral Coefficients (MFCC) after applying an extended spectral subtraction method, and Sphinx embedded denoising features from recent sphinx versions. Although removing C0 in MFCC-based features leads to a slight decrease in speech recognition performance, it makes the speech recognition system independent on the speech signal level. With multi-condition training, the three sets of noise-robust features lead to a rather similar behavior of performance with respect to SNR and presence of non-speech data. Overall, best performance is achieved with the extended spectral subtraction approach. Also, the performance of the PNCC features appears to be dependent on the initialization of the normalization factor.