How the speaking score works
The score is 0–100. Six components, fixed weights, a straight line between "good" and "bad" for each one. Nothing is adjusted per person, and the same formula runs on the website and in the app.
The formula
score = round(Σ(part × weight) / Σ(weight))
Each part is 0–100. If the language model fails, its parts are removed from both sums — the remaining weights are rescaled rather than replaced with a guess.
Fillers
30%Count of um, umm, uh, uhh, uhm, er, erm, ah, hmm, mm, mhm per minute of speech.
- 100:
- 1 or fewer per minute scores 100.
- 0:
- 10 or more per minute scores 0.
Pace
20%Words per minute, with fillers excluded.
- 100:
- 130–160 words per minute scores 100.
- 0:
- 80 or below, or 220 and above, scores 0.
Hedging
15%“I think”, “maybe”, “just”, “kind of”, “sort of” per minute.
- 100:
- 1 or fewer per minute scores 100.
- 0:
- 8 or more per minute scores 0.
Pauses
10%Silences longer than 1.5 seconds between words, per minute.
- 100:
- None scores 100.
- 0:
- 4 or more per minute scores 0.
Structure
15%A language model reads the transcript: is the point first, then support, then a close?
- 100:
- 90–100 means point first, supported, closed.
- 0:
- 50 means some order, but the point is buried.
Conciseness
10%A language model reads the transcript: repeats, detours, padding.
- 100:
- 90–100 means tight with nothing to cut.
- 0:
- 50 means about a third could be cut.
What happens to your audio
The browser converts the recording to 16 kHz mono WAV and uploads it once. The server sends it to a transcription provider that keeps filler words, computes every number, writes the feedback, and discards the audio. Nothing is written to disk, storage or logs. The transcript and scores are saved so you can see progress; the recording is not.
AssemblyAI transcripts are deleted through their API immediately after use, and Deepgram is called with training opt-out. The coach only ever sees text — never the audio.
Why these thresholds
They are deliberately conservative and public. A filler threshold of one per minute is close to scripted speech; ten per minute is where listeners start counting instead of listening. The pace band of 130–160 words per minute is the range that stays comfortable on a compressed video call. If your own experience says otherwise, the thresholds are still the ones being applied — the number should be predictable, not flattering.
What the number cannot tell you
Whether the idea was good. Whether the room was hostile. Whether the question was unfair. A score of 60 on a hard question may be a better performance than 90 on an easy one. Use it to find the next thing to practise, not as a verdict on the answer.
Questions
- Why is the transcript needed for scoring?
- Fillers, pace, pauses and hedging come from word timings. Structure and conciseness need the content. Both are measured on the same recording, so the number matches what you actually said.
- What happens if the language model fails?
- Structure and conciseness are dropped and the other four weights are rescaled to 100. The score still reflects what was measured, and the result says so.
- Does the score judge my accent?
- No. Accent is not an input. Pronunciation mode checks whether words were recognised, word by word, and never rates how you sound.
- Why does a short answer get no score?
- Under 15 words there is not enough speech for the pace and filler rates to mean anything.