VEO 3 Failed the Bawstin Test

Anyone who has spent time in Boston knows the accent is not just about dropping the letter R. It is a whole rhythm: "cah" instead of "car," "regulah" instead of "regular," a rotary instead of a traffic circle. So what happens when you hand an AI video generator a script written entirely in that dialect and ask it to perform it?

The prompt

The test used a short spoken-word script split between a female voice and a male voice, written phonetically to force the accent: "Welcome ta Bawstin," "the chowdah's hot," "you staht in the cah." The script leaned on the specific vocabulary that marks someone as a local rather than someone doing an impression of a local, bang a uey, Dunks, a regulah with cream and two sugahs. These are not generic "Boston accent" cues pulled from a stereotype; they are the actual shorthand locals use, which makes them a harder target for a model to reproduce convincingly.

Why this is a real test

Getting an AI to drop an R is easy. Getting it to drop the right R's, in the right words, while keeping the vowel shifts and cadence that make the accent recognizable rather than cartoonish, is much harder. Phonetic spelling in a prompt only goes so far when the underlying voice model has its own assumptions about how English sounds. The gap between "sounds like it's trying to do a Boston accent" and "sounds like it's from Boston" is exactly where a lot of AI voice generation still struggles.

Key takeaways

• Regional accents are a useful stress test for AI voice and video generation, because they demand precise phonetic and rhythmic accuracy, not just a general impression.
• Phonetic spelling in a prompt is a common workaround, but it does not guarantee the model reproduces the intended sound correctly.
• Vocabulary specificity (Dunks, a regulah, bang a uey) is as important as pronunciation for an accent to read as authentic rather than performed.

Try it yourself

If you work with AI video or voice tools, try writing a short script in your own regional dialect, phonetic spelling included, and see how closely the output matches what you'd actually hear on the street.