Keyword Co-articulatory Boundary Segmentation of the CRM Speech Corpus

Abstract

crmSegmentCallColorDigitNow_v1 contains segmentation data for the CRM speech corpus [R. Bolia, W. Nelson, M. Ericson, and B. Simpson, “A speech corpus for multitalker communications research,” Journal of the Acoustical Society of America, vol. 107, pp. 1065-1066, 2000.] Each line of the file corresponds to one of the corpus utterances and contains the ID of the Talker, the audio filename, the sample frequency, and the start and stop sample indices for the Callsign, Color, Digit, and Now words. For example, the fourth line of the data

Talker: 0, Filename: 000003, SampleFreq: 44100, CallStart: 10557, CallStop: 26975, ColorStart: 40317, ColorStop: 48910, DigitStart: 52218, DigitStop: 63916, NowStart: 67223, NowStop: 79767

lists the samples indices for CallStart (10557), CallStop (26975), ..., NowStop (79767) for the file 000003.wav from Talker 0, where the first two elements in the filename indicate Callsign (0-7), the next two elements indicate Color (0-3), and the final two elements indicate Digit (0-7).

The results were obtained by pre-segmenting the data using a combination of acoustic features. Fine-tuning of the segmentation was performed by a single listener (the author, G. Wakefield) with particular emphasis placed on minimizing the co-articulatory "leakage" at the segment boundaries. It should be emphasized that these particular segment boundaries, particularly for the word offsets, are often judgment calls and are typically based on whether the following phoneme can be predicted on the basis of the current segmentation.