Does AI transcribe every accent as General American? Variety normalization in phonetic transcription by multimodal models

Multimodal language models transcribe speech phonetically and are alreadyproposed for pronunciation teaching, with the transcription treated as a neutral record of thesignal. This study tests whether such a model preserves the variation between speakers ormoves it toward General American (GA). Gemini 3.1 Flash-Lite transcribed 117 SpeechAccent Archive speakers in five prestige tiers, three times each, under a standard and an anticitation instruction, and the study compared each output with the archive's expert IPA. Themodel's transcriptions of different speakers were separated by about a fifth of the distancebetween the expert's transcriptions of the same speakers (convergence ratio 0.20, 95% CI[0.18, 0.22]), by a similar factor in every tier, and lay closer than the expert's to a GAreference and, above all, to the citation forms of isolated words, with American rhoticity. Themodel recorded none of the vowel features that marked Australian, New Zealand andSouthern US speech, and gave /r/ to non-rhotic speakers in 73% of contexts, against theexpert's 5%. Neither the anti-citation instruction nor a minimal prompt brought thetranscription back to the speaker. A larger model, tested on 25 speakers, recorded more localfeatures but also features the speakers did not produce. The article argues that standardlanguage ideology enters the transcription through the reference the model applies to everyspeaker, and that AI transcriptions used in pronunciation teaching need to be checked bylistening. Claude (Anthropic) and Consensus assisted with literature search, analysis scripts,revision and translation.