Method
Everything below is published deliberately. A diagnostic you cannot inspect is an opinion with a number attached.
Where the numbers come from
Three measurement layers feed one deterministic scorer. No language model produces a score.
- Acoustic analysis (Praat) — pitch span in semitones, rhythm (nPVI), articulation rate, pause structure, voice quality, energy decay.
- Pronunciation model — phoneme, syllable and word accuracy, fluency and prosody, plus what you actually produced in place of the target sound.
- Verbatim transcription — disfluency counting. The transcript deliberately keeps every “um”.
A language model writes the prose only. It never receives your audio, and its output is checked in code: any figure not present in the measurements is rejected and the report falls back to a template.
Evidence gating
Each sub-criterion has a task that provides its primary evidence. If that task was not completed, or its recording failed the audio quality check, the item is reported not assessed with the reason. It is never estimated from something else.
| Sub-criterion | Primary evidence | Scale |
|---|---|---|
| Phonemic Accuracy | T1 | % of words mispronounced |
| Word Stress | T1 | % of words with misplaced stress |
| Articulation | T1 | score |
| Connected Speech | T1 | observational |
| Stress-Timing | T1 | observational |
| Pitch Variation | T2 | semitones |
| Terminal Contours | T2 | % correct |
| Uptalk | T2 | % declaratives rising |
| Focus Words | T3 | score |
| Contrastive Emphasis | T3 | prominence ratio |
| Speech Rate | T4 | syl/s |
| Disfluency Rate | T4 | per 100 words |
| Pause Placement | T4 | score |
| Vocal Stability | T4 | HNR dB (held vowel) |
| Projection | T5 | observational |
| Register Shift | T5 | observational |
Why five tasks
T1 — The Mirror
Every English sound appears in this paragraph. That is not an accident — it is the only way to check all of them in under a minute.
T2 — The Exchange
Questions go up, statements go down. Getting this backwards is the most common reason capable speakers are misread as uncertain — and it cannot be seen in a paragraph with no questions in it.
T3 — The Emphasis Drill
Same words, four meanings. Because the sounds are identical across all four readings, any difference we measure is purely emphasis — the cleanest possible test.
T4 — Off the Page
Reading hides things. It suppresses hesitation and hands you someone else's rhythm. When you choose your own words, your real fluency shows up — and that is what listeners actually hear.
T5 — Two Rooms
Sounding right at work and sounding right with friends are different skills. Register can only be measured as a difference, so we hold the news constant and change who is listening.
Eight of the items we report cannot be measured from read speech at all — disfluency, pause placement and register among them. Reading suppresses hesitation and hands you someone else’s rhythm. That is why T4 takes the script away.
Limits we are not hiding
- Bands are provisional. Boundaries are informed placeholders until a reference corpus and a test–retest reliability study are complete. Rubric version provisional-0.1.
- Projection is confounded by hardware. Microphone and gain affect it, so it is read within your own recording and never compared between people.
- Confidence is an interpretation, not a measurement. We present it as a composite of measured prosodic features, with its inputs shown.
- Accent is not the target. We report what costs a listener effort. A strong accent that costs nothing is not a finding.
Your recording
Audio is encrypted at rest and deleted 30 days after your assessment completes. The measurements derived from it are kept so a future re-test can be compared against this one.