Speech enhancement using phoneme level classification

Speech enhancement is one of the speech processing techniques which is perturbed mostly with increasing some perceptual part of speech like quality and intelligibility that has been debased by noise. The intelligibility of a speech can be improved by the isolation of target speech from noisy speech. This work aims to enhance the performance of the speech at the phoneme level. A phoneme is the basic unit of sound. There are 59 phonemes as per the TIMIT speech database. Features such as MFCC, RASTAPLP, AMS, GTFB are extracted for each phoneme at the frame level. The extracted features are given to the classifier for training as a result, a trained model is obtained. Then, Binary Mask based on the frame energy is estimated for each phoneme. An Ideal Binary Mask (IBM) is generated, if the energy of speech is greater than noise then it is assigned 1, otherwise, it is 0. During testing, the trained model gives the label for the given phoneme and the corresponding ideal binary mask is taken for resynthesis. Thus, enhanced speech is obtained. Performance such quality measure (SNR), intelligibility measure (STOI & CSII) were calculated for the resultant speech signal.