Published by TonePerfect. Facts about other providers come from their public pages, checked on 14 November 2026 (sources at the end). We have not benchmarked either of them.
Two different questions
Accent identification answers “which accent does this speaker have?” — a label such as American, British or Indian, usually with a probability. Pronunciation assessment answers “how close is each sound to a reference accent?” — scores per word and phoneme against a known text. Pronunciation APIs, ours included, do the second; they do not name the speaker’s accent. If you need a label, you need an accent-identification model.
Hosted API: Modulate accent detection
Modulate documents an accent detection endpoint (POST https://platform.modulate.ai/api/velma-2-accent-batch, key in an X-API-Key header) that classifies the regional or national accent from the voice signal and returns a whole-file label plus labels for successive time windows; it produces no transcript. Listed labels: American, British, Australian, Southern, Indian, Irish, Scottish, Eastern European, African, Asian, Latin American, Middle Eastern and Unknown. Files up to 100 MB in common audio formats. Its documentation page does not state a price.
Self-hosted: CommonAccent (open source)
The CommonAccent model on Hugging Face (Jzuluaga/accent-id-commonaccent_xlsr-en-english, MIT licence) is a fine-tuned XLSR model built with SpeechBrain that classifies 16 English accents, including us, england, australia, indian, canada, scotland, ireland, newzealand and others. Its model card says no hosted inference API is offered, so you run it on your own server and wrap it in your own endpoint.
TonePerfect: an accent guesser demo, not an API
Our learner site has a free accent guesser (beta): you read a 97-word passage, about 40 seconds, and a Whisper-based model estimates how likely English is your first language, then names the closest native variety (such as North American, English, Scottish, Irish or Australian) or first-language background (such as Indian, Mandarin or Spanish), with calibrated percentages, a language family when two accents are tied, and “hard to place” when nothing fits. It needs at least 15 seconds of speech, and we do not publish an accuracy figure for it yet.
It is not available through the TonePerfect API: there is no key-authenticated accent endpoint, and the demo’s backend only answers the toneperfect.app website. If you would like accent identification as an API, email us with your use case and volume.
What the TonePerfect API does instead
The pronunciation scoring API scores a recording of a known sentence sound by sound: for English (en-US, scored against an American reference) every word comes back with the expected and heard phoneme and a 0–100 score, so you can show a learner which sounds give their accent away, not which country it suggests. Ten languages, 50 free credits on sign-up with no card, then $10/month for 2,500 recordings. Code: add pronunciation scoring to a Python or Node.js app.
Before you use an accent label
- Test on your own users. Run recordings from the accents you actually serve, on the microphones they use; short or noisy clips are harder for every model.
- Treat it as a guess. Accents overlap and many first languages share features; show probabilities and allow “unknown”.
- Do not use it to decide things about people. An accent label is not a nationality, an origin or a measure of ability.
Sources checked 14 November 2026: Modulate accent detection docs · CommonAccent model card · TonePerfect accent guesser.