Distance Measures for Wavelet Representation of Speech Segments

Dyadic scheme of wavelet signal decomposition leads to a specific division of frequency bands. It is comparable to mel-frequency division and may be used in effective parameterization of speech signal in recognition systems, speech coding or other speech signal based applications. This paper discusses efficiency of different spectral distance measures applied to wavelet-parameterized speech. The presented methods are designated to use in isolated phoneme recognition task. INTRODUCTION Human ear is sensitive to frequency attributes of audio signals, including speech. Frequency resolution of human’s perception is not the same for all frequencies. In low frequency bands, slightly different tones are much more distinguishable then they would be in higher frequency bands. In the effect, frequency characteristics of human’s ear are not linear but rather logarithmic. Experiments showed that human aural system perceive frequencies not linearly but according to a mel-frequency scale [9] ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ + ⋅ = 700 1 1127 Hz e mel f f log . (1) Many practical works confirmed usefulness of this scale in speech recognition systems, speech coding or compression. Most frequently used type of speech parameterization that support melfrequency scale is MFCC (mel-frequency cepstral coefficients) [6]. Our proposition provides logarithmic frequency scale division, typical for dyadic discrete wavelet decomposition scheme, and may be an interesting alternative approach. It is necessary to adjust known, or develop new computational tools for use with this type of parameterization in speech recognition applications. As far as wavelet parameterization scheme, has properties similar to linear prediction coefficients or mel-frequency cepstral coefficients, the known computational tools may be applied. In this paper different spectral distortion measures were applied to wavelet parameterization scheme in order to determine their efficiency. WAVELET SPEECH SIGNAL PARAMETERISATION Continuous-time wavelet transform is defined as