A speech template matching technique based on subspace approach

A speech template matching technique based on subspace approach is described. First, the normalized speech waveforms are obtained from the training data by a suitable time-axis warping transformation under the assumption of a statistical dependency of the time sequence of the observation vectors. A basis is constructed from the n_i eigenvectors corresponding to the eigenvalues approaching zero and it is used to obtain a projection operator for each class. The projection operators are then used to classify the unknown pattern into the class on whose class subspace it has the longest projection. The technique is tested on a set of experiments constructed using nearly 8000 utterances of the 26 letters of the British alphabet spoken by 104 speakers, roughly half of which are male and the other female. The best classification rate of 77.8% is obtained from the experiment which is carried out using the letters “b, d, g”