A study of multilingual speech features: perceptive scalogram based on wavelet analysis

In recent years adopted speech feature for automatic speech recognition system is degraded by the additive noise when the system is being used in real environments. In this respect a lot of efforts have been devoting to find the robust features. In this paper, we propose the use of perceptive scalogram, which is a kind of time-scale distribution of signal to approximate perception of speech. It provides a three-dimensional representation of speech signal spectra, with time represented on the x-axis and frequency on the y-axis. Gray scales are used to represent the third dimension, either the amplitudes (for scalograms) or the phase (for phase-shift diagrams), of the wavelet transform data. It takes considerations that speech and background noise has different perceptions on human beings, and speech signals are non-stationary. It is a time-scale distribution of signal, and uses wavelets as its analysis tool, which makes a good approximation of some hearing properties. Computer simulations show that the proposed perceptive scalogram is very robust to additive white Gaussian noise for different languages (e.g. English, Putonghua, Cantonese)