Frequency bark cepstral coefficients extraction for speech analysis by synthesis.

Abstract

The Mel scale was proposed in 1937, following a series of experiments used to establish a perceptual scale. The use of the Mel scale is almost standard for speech recognition application. The Bark scale divides the audible spectrum into 24 critical bands that try to mimic the frequency response of the human ear. This article describes the process used to extract a set of cepstral coefficients from a warped frequency space (Mel and Bark) and analyze the perceived differences in the reconstructed signal. We will try to determine if there is any audible improvement between these two most warping functions for the purpose of speech analysis by synthesis. We will use the same procedure for parameter extraction and signal reconstruction for both functions, replacing only the warping scale used, to minimize the distortion other elements might add to the results. After running the waveform through both processes and reconstructing a wave signal from the parameters, while the resulting output was somewhat different, there were slight differences between the bark and mel generated signals. Statistical tests are now running between Mel and Bark scales.