PhonoPerfect

AbstractThis study introduces a comprehensive phoneme-based dataset designed for training and evaluating models in mispronunciation detection and correction. The dataset, derived from publicly available speech resources, comprises approximately 275,000 rows of words and their corresponding phonemic transcriptions in the International Phonetic Alphabet (IPA) format. To meet the requirements of both correct and mispronounced phoneme data, three transformation methods—substitution, insertion, and deletion—were systematically applied to generate mispronunciations. These transformations were guided by the Acoustical and Articulatory Phonic Similarity Act, ensuring realistic emulation of pronunciation errors observed in actual speech. Substitutions were based on phonemic similarity, with consonants replaced by others sharing similar acoustic properties or articulation (e.g., plosives /p/ ↔ /b/, nasals /m/ ↔ /n/), while vowels were substituted by those with proximal tongue height or rounding (e.g., /i/ ↔ /ɪ/). Diphthongs were replaced with alternate vowel transitions common in other languages, and tonal inflections were adjusted to mimic pitch errors in tonal languages. The dataset, consisting of words, their correct pronunciations, and systematically generated mispronunciations, provides a robust foundation for training and validating phoneme-level detection and correction systems. This resource is particularly suited for advancing applications in language learning and speech therapy.