Analysis of hesitations in spontaneous speech

Abstract

Spontaneous speech differs from read speech in several ways, especially in hesitation phenomena. This paper reports results on hesitation pauses (filled and unfilled) and restarts. For comparison purposes, the acoustic correlates of (unintended) hesitation pauses are compared to those for intentional pauses. A distinction is made between grammatical pauses (at major syntactic boundaries) and ungrammatical ones. Such pause types cannot be separated based on silence or prepausal duration, but rather in the pitch of the prepausal word. Ungrammatical pauses tended to have few F0 continuation rises, whereas virtually all grammatical pauses were accompanied by a prior F0 rise of at least 10 Hz. While silent pauses are easy to locate in speech recognition applications, filled pauses (e.g., “err,” “umm”) resemble words in continuous speech. Filled pauses at major syntactic boundaries were about 300–450 ms, whereas those within syntactic units were shorter. Filled pauses had falling or flat and low F0 patterns. Ones at syntactic boundaries tended to start higher in F0 and then fall, whereas filled pauses internal to a syntactic unit had lower F0 patterns. Concerning restarts in spontaneous speech, when a work was completely repeated, it had virtually the same prosodics in both its instances. When a word was changed in the restart, its second instance was more stressed. [Work supported by Canadian government.]