Output details

Analysis and prediction of acoustic speech features from mel-frequency cepstral coefficients in distributed speech recognition architectures

<22>It was widely thought that MFCC features used in speech recognition retain only the ‘spectral envelope’ and are free from source information. Correlation analysis presented in this work dispelled that thinking and demonstrated that MFCC features also include voicing and fundamental frequency information. This led to a method of estimating acoustic speech features from MFCC features (as encountered within distributed speech recognition) and ultimately speech reconstruction, which was not considered possible solely from MFCC vectors. This work formed part of a wider project into speech enhancement (EPSRC GR/S30238/01) with Brunel University and the Institute of Sound and Vibration Research.