How long does an AI need to listen before it can place your accent? It's the kind of question that sounds philosophical but has an exact, measurable answer — and we measured it, because our first instinct was wrong in both directions.
We took thousands of recordings of English speakers from around the world and asked the same model the same question on progressively shorter versions of each clip: the full recording, then twenty-five seconds, twenty, fifteen, ten, five, three. Same voices, same test, one variable — how much of the voice the model was allowed to hear. To keep the answer honest, the speakers used to measure accuracy were never the speakers the system learned from.
Three seconds of speech is nearly useless — barely better than guessing for the hard distinctions. A few seconds simply can't contain enough of the sounds that give an accent away. From there the curve climbs steeply: every added second exposes new vowels, new consonant transitions, more of the speaker's rhythm. Broad, regional judgments stabilize first. The finer judgment — which accent within a family — keeps improving noticeably longer, and stops improving at roughly twenty seconds of actual speech. Past that point, more audio adds essentially nothing: the model has heard your phonetic repertoire once, and hearing it twice tells it nothing new.
An accent isn't one sound — it's a distribution of habits across all the sounds of a language. A short clip is a random sample from that distribution, and small samples miss things: maybe the vowel that would have given you away just never occurred. Twenty seconds of a well-chosen passage is about the point where an English speaker has produced most of the sound inventory that matters. After that, the evidence repeats rather than accumulates.
Two design decisions fell straight out of the curve. First, our recorder runs for twenty seconds — not because users are impatient (they are), but because that's where measurement says the information stops arriving. Asking for more would cost patience and buy nothing. Second, everyone reads the same short passage — same words, same conditions — so the comparison between accents is fair and every second of the window counts. The seconds are few; we make each one carry evidence.
Twenty seconds. That's the whole price of finding out how you sound.
take the 20-second accent test →