Wavelet parameterization for speech recognition: variations in translation and scale parameters

This paper describes two experiments to isolate the processes of using the wavelet parameterisation with a hidden Markov model (HMM) for speech recognition. Previous work with the wavelet transform and the HMM highlighted a local maximum in recognition performance for varying frame advance and initial wavelet filter length. The first experiment varies the initial wavelet filter length (frequency resolution) with a fixed frame advance (fixed time sampling). This shows a local maximum with increasing initial wavelet filter length (increasing frequency resolution). The second experiment varies the frame advance (time sampling) with a fixed initial wavelet filter length (fixed frequency resolution). This shows recognition performance decreasing monotonically with increasing the frame advance