Interaction of acoustic and semantic context information on phonetic identification

Background : Speech comprehension depends on both phonetic and semantic processes. Phonetic mechanisms aim at categorizing phonemes from the acoustic input. Various sources of information are available among which some are intrinsically related to the current segment (formant frequencies, formation ratios) while others relate to extrinsic information relating to acoustic context (e.g. speaker-specific properties associated with neighbouring acoustic information). These processes would interact with a semantic analysis of the message in order to provide an optimal interpretation of the signal. The influence of semantic information on phonetic identification is also viewed as a context effect. Therefore, two different sources of context information are available to listeners : phonetic context is associated with how listeners adapt to acoustic information neighbouring the target segment, while semantic context is associated with how the identification of neighbouring words along with sentence interpretation affect speech perception. Both context mechanisms may interact with bottom-up acoustic analysis in order to reach the final categorisation.
Our aim is to investigate how these two sources of contextual information interact with each other in speech perception. More precisely, we want to assess how the acoustic information is processed and related to contextual information when acoustic and semantic contexts interact together.

Method : In order to address this issue, we designed French linguistic material in which a target word is preceded by a subject-verb structure which is expected to favour one interpretation over the other. Target words were selected in order to constitute CV or CVC word pairs exhibiting vowel alternations associated with distinct but acoustically close vowels (e.g. « balle » (/bal/) / « belle » (/bɛl/)). Three semantic contexts were designed for each word pair (Context 1 favours Word 1, Context 2 favours Word 2, Context 0, neutral, favours none). These semantic contexts are combined with synthetic modifications of the formant frequencies in the first part of the carrier sentence (excluding the target word). The final word is selected among a set of words resynthesized along a 7-step formant continuum (F1/F2/F3/F4) in order to provide measurements to compute psychometric curves. 2-AFC classification frequencies and psychometric modelling are computed in order to investigate how acoustic and semantic information interact. Our predictions are that acoustic contextual information will compete with semantic information. The results will then provide the basis for modelling uncertainty processing in speech perception. Data collection is still in preparation and preliminary results will be presented during the conference.