Methods for phonemic recognition in speech processing.

Abstract

Speech is one of the most inefficient methods of communication. Therefore, there has been a continuing effort to devise means to reduce the redundancy, that is, compress the bandwidth required for speech communication „ For broad tactical military use the author considers the amount of compression, intelligibility and quality to be factors of prime importance, while factors such as speaker recognition and naturalness are of secondary importance The speech compression methods receiving the most emphasis today are described and their major discrepancies indicated . Generally, their deficiencies arise because the present systems do not rely on the fact that the electrical representation of speech is a particular signal, not just any electrical signal whose frequency components lie in the audio band. With present day speech compression systems in mind, an analysis of the "method of distinctive features" as proposed by Jakobson, Fant, and Halle of MIT is offered This method for achieving reliable speech recognition at the phonemic level is in a partial stage of development » The measurements required to extract six of the ten features are tabulated and procedures for reducing the remaining four features are outlined, Sonagraphic data is given in support of the method. Instrumentation and flow charts for digital computer investigation of the process are detailed. The proposed acoustical input device will facilitate computer studies of speech processes.