Toward Objective Phonetic Transcription - An On-Line Interactive Technique for Machine-Processed Speech Data

The derivation of phonetic transcription is one component of speech processing that can utilize man-data techniques. This involves the assignment and time location of phonetic symbols to speech data. Conventional transcription methods are suspect in at least two respects. First, the symbols assigned are often based on "talker intention," or what "should be," rather than on physical evidence; and, second, symbol location is usually related to subsequent machine processing only minimally. An exploratory procedure is described that operates on-line, in an interactive mode, utilizing a graphic display with an IBM System/360 model 40. Available for display are a digital sound spectrogram, the power spectrum at a given time sample, an average spectrum for a given sound class, and correlations of a current spectrum with average spectra. Using this technique, location and assignment of phonetic symbols are less dependent upon operator expertise (and/or bias) and highly related to subsequent processing.