A microscopic analysis of consistent word misperceptions.

Abstract

Speech misperceptions have the potential to help us understand the mechanisms involved in human speech processing. Consistent misperceptions are especially helpful in this regard, eliminating the variability stemming from individual differences, which in turn, makes it easier to analyse confusion patterns at higher levels of speech inits such as the word. In this thesis, we haver a conducter an analysis of consistens word misperceptions from a "microscopic" perspective. Starting with a large-scale elicitation experiment, we collected over 3200 consistent misperceptions from over 170 listeners. We investigated the obtained misperceptions from signal-idependent and a signal-dependent perspective. In the former, we have analysed error trends between the target and misperceived words across multiple levels of speech units. We have shown that the error patterns observed are highly dependent on the eliciting masker type and contrasted our results to previous findings. In the latter, We attempted to explain misperceptions based on the underlying speech noise interaction. Using tools from automatic speech recognition, we have conducted an automatic classification of confusions based on their origin and quantified the role misallocation of speech fragments played in the generation of misperceptions. Finally, we introduced modifications to the original confusion eliciting stimuli to try to recover the original utterance by providing release from either themasker`s energetic or informational component. Listeners¿percepts were reevaluated in response to the modified stimuli which revealed the origin of many confusions regarding energetic or informational masking.