Automating Error Frequency Analysis via the Phonemic Edit Distance Ratio

Purpose Many communication disorders result in speech sound errors that listeners perceive as phonemic errors. Unfortunately, manual methods for calculating phonemic error frequency are prohibitively time consuming to use in large-scale research and busy clinical settings. The purpose of this study was to validate an automated analysis based on a string metric-the unweighted Levenshtein edit distance-to express phonemic error frequency after left hemisphere stroke. Method Audio-recorded speech samples from 65 speakers who repeated single words after a clinician were transcribed phonetically. By comparing these transcriptions to the target, we calculated the percent segments with a combination of phonemic substitutions, additions, and omissions and derived the phonemic edit distance ratio, which theoretically corresponds to percent segments with these phonemic errors. Results Convergent validity between the manually calculated error frequency and the automated edit distance ratio was excellent, as demonstrated by nearly perfect correlations and negligible mean differences. The results were replicated across 2 speech samples and 2 computation applications. Conclusions The phonemic edit distance ratio is well suited to estimate phonemic error rate and proves itself for immediate application to research and clinical practice. It can be calculated from any paired strings of transcription symbols and requires no additional steps, algorithms, or judgment concerning alignment between target and production. We recommend it as a valid, simple, and efficient substitute for manual calculation of error frequency.