Why AI speech scores vary between takes
AI speech scores vary between takes because the report is written by a language model that is allowed to vary its wording and marks slightly from run to run. So the same answer can score a little differently when you record it again. The transcript, the lighting and the length of the take also change the evidence the report works from. Speech-to-text depends on how clearly the words were captured, and face tracking depends on the camera picture. A short take of thirty seconds gives the report fewer moments to judge than a take of three minutes. A single score is one sample, not a fixed measurement, so read the key moments and the transcript before the number. To see your own spread, record the same answer twice in the same light and at the same distance. We have not yet published a measured figure for how much scores vary between repeat takes.
The report is written by a language model that is set to vary its wording and marks slightly from run to run, so the same answer can score a little differently when you record it again. A single score is one sample, not a fixed measurement. The transcript, the lighting and the length of the take also change the evidence the report works from. Read the key moments and the transcript before you read the number. We have not yet published a measured figure for how much scores vary between repeat takes.
What causes the variation
The scoring model is allowed to vary
The model that writes the report does not give the same wording every time. Its setting leaves room for variation, so the marks and comments can move a little between runs, even on identical input.
The transcript changes with the audio
Speech-to-text depends on how clearly the words were captured. Background noise, a quiet microphone or a fast delivery can change which words the report sees, and the feedback follows the transcript.
Face tracking depends on the picture
Expression signals are measured from the camera. Low light, a camera below eye level or a face partly out of frame give the analysis less to work with.
A short take has less evidence
Thirty seconds of speech gives the report fewer moments to judge than three minutes. Short takes are more likely to swing, and we would read them with more caution.
How to read a score
- Look at the key moments first. Each one points to a time in your recording, so you can check it yourself.
- Read the transcript against the feedback. If a point refers to something you did not say, the transcript is the place to check.
- Record the same answer twice in the same light and at the same distance. The gap between the two shows your own spread, which is more useful than any single number.
- Treat the coaching tip as the action to practise, whatever the score was.
What we have not measured
We have not yet published how much scores move across repeat takes of the same answer. Until we do, use the score to compare evidence and to spot patterns over several sessions, not to judge a single take as a fixed result.
Record the same answer twice
Ten sessions for $12.99, paid once, so you can compare takes for yourself.
Get 10 sessions - $12.99