First International Conference on Spoken Language Processing (ICSLP 90)
In this paper we present a system designed for the broad labelling of a large speech acoustic database. This system uses an original labelling strategy which is based on 3 characteristics: a time-expanded (10 times) speech sound output, an automatic scanning of the corresponding label strings and a speech visualization (using a moving cursor which is synchronized with the time-expanded speech sound output and the label strings scanning). By using an algorithm that is similar to the SOLA algorithm proposed by Roucos , the time-expansion technique produces a good quality 10 times expanded speech signal from the speech signal to be labelled. The time-expanded speech listening brings out a fine speech structure. This result is usually attained, with difficulty, by careful spectrogram reading in a classic manual labelling system (signal editors).The advantage of the automatic label strings scanning mechanism is that it enables the cursor to be moved by the system instead of by the operator. Eventually the operator labelling task is only confirmation of the current label. Moreover the time-expanded mechanism can be viewed as a kind of " acoustic magnifying glass " which brings some additional perceptual information on the signal.
Bibliographic reference. Tuffelli, D. / Wang, H. D. (1990): "TELS: a speech time-expansion labelling system", In ICSLP-1990, 1001-1004.