Showing papers on "Speaker diarisation published in 1973"

Effects of differing phonetic contexts on spectrographic speaker identification.

Abstract: This study investigated how well the file card system of “Voiceprint Identification” reported by Kersta fulfills its purpose of minimizing the effects of contextually caused spectral variation, and how well it serves as either an identification or population reduction tool. Subjects received training believed to be equivalent to that received by the experimenter in the Voiceprint Identification Training Course, and achieved error rates approximating those reported by Kersta for similar training tasks. Experimental tasks required subjects to identify unknown speakers from a population of 50 known speakers by first excluding all known speakers they were certain were not the unknown speaker, and then attempting absolute identification or elimination. Attempts were made under five experimental conditions created by combining two variables, phonetic context and inclusion of the unknown speaker in the known speaker population. The data show that the system tested does not effectively reduce the effects of contextual variation, and cannot be used for either absolute identification or elimination, or population reduction.The data suggest that the value of spectrograms for speaker identification purposes is limited to use as an investigative aid, and then only if speech samples of similar context and adequate duration are compared.

Techniques for Objective Speaker Identification

Abstract: A technique has been developed and tested for identifying from voice recordings which, if any, of a number of possible speakers corresponds to a particular object speaker recording. Segments of the same set of selected phonemes by the object speaker and all suspect speakers are Fourier transformed. These phoneme spectra, and features derived therefrom, are utilized to make measures of voice similarity. Experiments with voice recordings by 140 men have yielded estimates of the probability of correct identification as a function of the number of suspects for both forced‐choice and open‐choice tests. Identification reliability has also been related to number of phonemes, number of utterances, and number and type of features. Open‐choice decision reliability of 99% or better for small numbers of suspects was demonstrated. Another decision procedure, whereby decisions are made in only 95% to 99% of cases, can reduce error probabilities essentially to zero.