Publication Abstract

Jankowski, C., Kalyanswamy, A., Basson, S. and Spitz, J. NTIMIT: A phonetically balanced, continuous speech, telephone bandwidth speech database. Proc. ICASSP-90, pp. 109-112, 4/3-6, 1990.

Abstract

The creation of a continuous speech, multi-speaker, telephone bandwidth speech database is described. The NTIMIT (Network TIMIT) database was collected by transmitting the TIMIT database over the telephone network. Additional advantages of the NTIMIT database include a carefully selected diversity of speech dialects and extensive breadth and depth of phonetic coverage. NTIMIT is orthographically and phonetically labelled identically to the TIMIT data. Possible uses for NTIMIT include acoustic analysis of telephone bandwidth speech, development of telephone bandwidth speech recognition algorithms, and retraining current wideband algorithms for telephone speech. Speech transmission was achieved by creating a "loopback" telephone path to a large number of central offices. The central offices were geographically distributed to simulate different telephone network conditions. Half ot the TIMIT database was sent over "local" telephone paths, while half was transmitted over "long distance" conditions. Transmission involved the use of a commercial device to simulate the acoustic characteristics between a human's mouth and a telephone handset. All recordings were done in a acoustically isolated room. Calibration signals were transmitted to each central office in order to readily evaluate such network characteristics as attenuation, frequency response, and harmonic distortion.