MPO-APP with preprocessing for speech enhancement and noise robust ASR

We have developed a speech enhancement algorithm that is based on modified phase opponency (MPO) and a periodicity measure (Deshmukh et al. 2007) and does not need any estimate or statistical characterization of the noise.  The performance of the proposed enhancement scheme, evaluated using different objective measures, is comparable to that of some of the other speech enhancement schemes when the characteristics of the background noise are not fluctuating. However, the proposed MPO-APP enhancement scheme outperforms other speech enhancement schemes when the speech signals are corrupted by fluctuating noise.  Finally, unlike other enhancement schemes, the MPO-APP enhanced speech does not contain any musical noise and it is not reverberant.
Summary

Our preliminary research shows improvement in ASR experiments when noisy speech is enhanced by using the preprocessor based MPOAPP algorithm. The recognition rates are found to improve further when this algorithm is used in conjunction with variable frame rate analysis. The WER results for the car noise section of Aurora-2 database is shown below.
Text Box: Text Box: Text Box: Text Box: Clean Speech Text Box: Car Noise @ 5dB Text Box: MPOAPP processed with 20dB attenuation Text Box: Text Box: Preprocessor 2 + MPOAPP with 20dB Text Box: Text Box: Preprocessor 1 + MPOAPP with 20dB Text Box: Preprocessor 3 + MPOAPP with 20dB

Speech Enhancement