Perception of word boundaries in spontaneous speech

Abstract

Studies of lab speech have identified several acoustic variables that listeners use to identify word boundaries (e.g. allophonic variation, segment duration), however spontaneous speech may not contain the same acoustic signals. Using data from a recent study we tested whether disambiguating information is also available to listeners under more typical conditions. 15 pairs of target phrases only differing in the placement of word boundaries, e.g., grey#day vs. grade#A were elicited under read (146 tokens) and spontaneous speech conditions (316 tokens) from multiple talkers who were not aware of the ambiguities. Phrases were played in isolation to 30 Native English listeners in a 2AFC segmentation task (e.g. grey day vs. grade A). Accuracy was above chance for both read (80.9%) and spontaneous speech (73.1%). Accuracy varied according to the critical consonant type and was highest for voiceless stops followed by nasals, voiced stops, fricatives and clusters. Only two items (pierced ears vs. peer steers and beef eater vs. bee feeder) were at chance for spontaneous speech. Measures of word-initial lengthening were good predictors of performance but listeners outperformed predictions from these measures indicating they weren’t the only cues.