Transcript evidence
What the speech-recognition model heard in the recording.
Limit: Recognition errors, background noise, names, and code-switching can make the transcript wrong.
Scoring methodology
Practice evidence, with its limits visible
MyAccentApp gives pronunciation and communication practice feedback. Its scores are estimates for a single recording—not exam results, diagnoses, identity judgments, or proof that one accent is better than another.
What the number actually measures. The score is a recognition score: we transcribe your recording, then compare the transcript—word by word—against how your target accent’s dictionary says those words are pronounced. So it answers “was each word recognisable?”, together with timing and pitch. It does not measure how closely you realised the accent: if the model recognises the word, a very British and a very American rendering of it can score the same. Accent-feature detection is reported separately, and is a rough indication rather than a measurement. Scoring the accent itself needs a model that reads sounds directly from audio, which we are still building.
What a result can be based on
Timing evidence
Word timestamps supplied by the transcription service, used for pace, pauses, and synchronized markers.
Limit: Timing is omitted when the service does not return grounded word spans. It is never reconstructed from text alone.
Acoustic evidence
Measurements such as pitch range, intensity variation, pause ratio, and duration produced by the phoneme service.
Limit: These measurements describe this recording, not a speaker's identity, fluency level, health, or overall ability.
Confidence and abstention
Each attempt is labelled high, medium, or low confidence from the evidence available for that attempt. Low-confidence attempts hide the numeric score and do not update learning history or mastery. Medium confidence is shown as a rough estimate. A retry prompt is an honest abstention, not a score of zero.
Common reasons include unavailable acoustic analysis, too little speech, uncertain alignment, silence, or noisy audio. The result screen exposes which evidence was available and the scorer version when reported.
What has—and has not—been validated
Deterministic fixtures and local evaluation tooling test contracts, evidence labelling, retry behavior, and metric calculations. A frozen, independently rated real-speaker benchmark is still pending. Until that work is complete, MyAccentApp does not claim certified accuracy, fairness across demographic or language groups, or equivalence to IELTS, TOEFL, clinical, or professional assessment.
The scorer has not been validated for children, atypical speech, or mixed-accent recordings. Learners should treat corrections as prompts to listen, compare, and practise—not as judgments about their identity.
Where search clips come from
- YouTube videos.We index the captions of public videos and play the matching moment in YouTube’s own player, with the creator named. We assign each video’s accent label when we add it; the label is our judgement, not a certification.
- Openly licensed speeches. Recordings such as Voice of America programmes are hosted only after their licence has been checked and recorded. Each speech page names its licence.
- LibriVox audiobooks, only if you opt in.Public-domain volunteer readings. Readers rarely state where they are from, so these clips are always labelled “accent not verified”.
- Synthesized references. When no real clip matches, you can ask for a synthesized voice. It is labelled as synthesized and is only generated when you press the button.
Recording and data controls
Audio retention is explicit and feature-specific. Communication Studio defaults to discarding uploaded audio; keeping it for 7 or 30 days needs separate consent and is not consent to train a model. Signed-in users can export or delete account data.