Accented and non-native speech breaks pronunciation dictionaries in ways that quietly poison your acoustic model alignments, and this Interspeech 2021 work asks a deceptively simple question: what if we just added the detected mispronunciations to the dictionary and retrained? Non-Collision Mispronunciation Addition (NCMA) is the technique behind PAII's submission to the Interspeech 2021 Non-Native Child English Close Track ASR Challenge, and it is aimed squarely at the WER pain that kids' speech, L2 learners, and accented adult speech impose on production ASR.

The presentation walks through how NCMA avoids collisions when injecting new pronunciations into the lexicon, how better phone alignments feed back into a stronger acoustic model, and how the whole detect-add-realign loop translates into measurable WER drops on a genuinely hard benchmark. The technique is grounded in the paper published at Interspeech 2021 by W. Chu, P. Chang, and J. Xiao, and the presentation ties the algorithmic choices back to the challenge submission that motivated them. If you are building ASR for education, kids' products, computer-assisted language learning, or any long tail of accented speakers where a generic dictionary quietly fails you, this is worth the time to see how targeted lexicon repair beats brute-force data collection.