EVALUATION OF SPEECH PROCESSING DEVICES I. INTELLIGIBILITY, QUALITY, SPEAKER RECOGNIZABILITY

摘要

This study is concerned with the development of improved methods of evaluating experimentally processed speech and, in turn, speech-processind devices and systems. Three bases of evaluation are dealt with in the study. These are: Intelligibility, Speaker Recognizability and Aesthetic Acceptability or Quality.nA two-choice diagnostic rhyme test for the transmission of con¬sonant information has been developed. It yields a total intelligibility score plus diagnostic scores relating to the fidelity with which seven binary attributes of consonant phonemes are transmitted to the ear of the listener. These attributes are voicing, nasality, duration and frication (as opposed to piosion) i.e., front (as opposed to middle) middle (as op-posed to back) and back (as opposed to front).nFor treating the problem of speaker recognizability, procedures have been developed by means of which listeners' ratings of voices on vari¬ous perceived acoustic traits can be analyzed to predict speaker recogniza-bility under any given transmission condition.nThe problem of evaluating the aesthetic acceptability or quality of transmitted speech is treated by means of the standard unit-variance method. Here, primary emphasis is placed upon the contributions of the channel to the quality of the received speech. However, the method is adaptable for purposes of studying qualitative variation attributable to the source (i.e., the speaker). In this method, speech as processed by four representative vocoder systems provides standards with which experimentally processed speech is compared by listeners. Listener response data are analyzed to yield a value representing the position of the experimental system on a standard unit-variance scale of aesthetic acceptability.nResults of evaluations of representative vocoders are presented for each of the three evaluation methods.