This study aims to predict mishearings in contexts where misperception can be overlooked due to their semantic similarity. First, we constructed a mishearing corpus by collecting 4940 instances from diverse web sources, including medical incident reports, shorthand records, casual conversations, and taxi dispatch logs. After removing duplicates and implausible cases, 1884 instances were retained. Next, a prediction model integrating lexical priming, word frequency, and phonetic similarity was developed. From a vocabulary set of 2000 words, contextually activated candidates were extracted, and phonetic similarity to the input word was evaluated. As a result, 280 cases (14.8%) were predictable from context, with a mean rank of 22.8 and a mode rank of 1, while the remaining 85.2% included casual speech where contextual prediction was difficult. Furthermore, to account for individual hearing differences, four participants with distinct audiograms (mean hearing level: 16.2 dB) completed a phoneme identification task to derive confusion matrices. By incorporating these matrices, phoneme-specific edit distances were adjusted for each hearing profile, enabling personalized mishearing predictions. The model predicted an average of 54.8 candidate words (maximum 194, minimum 12), reflecting individual auditory characteristics. This study highlights the value of combining contextual information and auditory profiles to improve mishearing prediction.
Kishiyama et al. (Wed,) studied this question.