A METHOD OF DECODING SPEECH

A method is presented for decoding speech that uses 'machine events' as the basic linguistic units, in contrast with phonemes and other units employed in other studies. The machine events are represented as multidimensional binary vectors. Utterances of words are expressed as sequences of binary vectors. Spoken words are decoded as sub-sequences of machine events. Decoding is independent of time, i.e., it does not depend on the duration of the speech signal. Results are given of small-scale tests of the decoding method. The experimental apparatus consisted of a 12-channel short-time spectrum generator and a machine-event generator that implemented six hyperplanes and a human who assumed the role of a translator in transcribing the six recorded outputs of the machine-event generator into sequences of six-digit binary number representations. Unambiguous decoding was obtained based on recordings of five utterances for each of ten words (the digits, 'zero' to 'nine') by each of four different speakers. The report discusses theoretical foundations for the decoder, design principles for the short-time spectrum and machine-event generators, and bases for the computation of significant sub-sequences from transitions of the machine events. Appendices describe the method for determining the distribution of the bandpass filters and for computing filter bandwidths. The method suggests a feasible approach to practical speech recognition devices. (Author) Contracts AF 33(615)-2573, AF 33(615)-3890, Grant PHS-GM-10718.