A Comparison of DFT, PLP and Cochleagram for Alphabet Recognition

The English alphabet is a small but dificult vocabulary fo r speech recognition, with many fine phonetic distinctions, such as M / N and B/V. We use speakerindependent classification of isolated English letters t o evaluate the relative performance of the D F T , Perceptual Linear Predictive analysis, and the cochleagram auditory model. Feedforward neural network classifiers were trained using all three representations o n 60 speakers and tested on 60 new speakers. Training and testing data were independently modified b y adding two levels of Gaussian noise and babble (20 random letter utterances, attenuated and given random offsets). PLP gave the best results, especially when trained or tested on Gaussian noise.