O-Engineers O-Engineers Aug 2017 | Page 46

Therefore, it suppresses the Spectral components that change more slowly or quickly than the typi- cal rate of change of speech. The RASTA approach can be combined with the PLP method to get the LP transfer function Hz [7]. Unlike cepstral mean subtraction, which removes the dc component of the short-term log spectrum, RASTA processing influences the speech spectrum in a more complex manner and emphasizes spectral transitions. The use of RASTA processing has been shown to im- prove speech-recognition performance under mis- matched environments. This band pass operation, combined with BPL filtering, has been shown to im- prove speaker-recognition performance under mis- matched conditions. However, this is not the end, currently speech rec- ognition communities are working on GMM and HMM, Long Short-term Memory RNNs based ap- proach as well as deep neural network which takes the speaker recognition to the next level and 46