Speech Sound Classification Based on Signal Statistics Free

A method of quantizing the output of a vocoder‐type speech analyzer is described. This quantization effectively classifies the sound at each instant into one of a small number (typically, less than 100) of classes, but in such a way as to allow relatively accurate reconstruction of the original unquantized voltages The boundaries of the classes are chosen on the basis of the statistical distribution of the voltages, using redundancies which may exist to reduce the number of classes to a low value. Experimental results are presented, illustrating classification schemes resulting from this procedure. These schemes bear some resemblance to conventional phonetic analyses of speech sounds, but they have the advantage of being derivable by purely mechanical means, without the use of human intuition or judgment. The use of such sound classification for automatic speech recognition is discussed.