How the numbers work
Every bubble shows two scores and a set of audio measurements. Here is exactly where each one comes from, so you know when to trust the translation and when to ask someone to repeat themselves.
Heard — transcription confidence
The speech model returns a probability for every token it emits. We take the geometric mean of those probabilities, which is length-normalised: a long confident sentence and a short confident phrase score the same, and one shaky word cannot be hidden by a hundred easy ones.
Tapping the meter also shows the least certain word and its probability. That single number is usually a better warning sign than the average: if the weakest token is a name or a number, treat it as unverified.
When the provider does not return token probabilities, the meter reads n/r — not reported. We never substitute a guess.
Language — candidate probabilities
Instead of asking a model how confident it feels, we ask for the two or three most likely languages with probabilities that form a distribution, then normalise them. The score on the bubble is the top candidate's probability; the runner-up is shown in the breakdown, which is what makes a 62% meaningful — you can see what it might be instead.
Short clips cannot support a confident answer, so the score is capped when there is under 1.5 seconds of audio or fewer than four words of evidence. The reason for the cap appears in the breakdown.
Audio quality — measured, not guessed
Length, average level, peak level, clipping percentage and a rough signal-to-noise estimate are computed directly from the captured PCM on your device. They explain low scores honestly: quiet audio, clipping from a voice too close to the mic, or a noisy room all degrade transcription in predictable ways.
Live drafts vs final bubbles
While someone is still talking, a dashed draft bubble streams in the words as they are recognised from a rolling window of the audio. Drafts intentionally carry no confidence scores — they are a preview. When the speaker pauses, the full clip is transcribed once more, detected, translated, and the draft is replaced by the final bubble with scores attached.
Each draft is a separate AI call, so they are throttled and can be turned off in settings if you want the cheapest possible run.
When to distrust a bubble
- Either meter under 35% — the bubble is drawn with a dashed edge.
- The runner-up language is within ~15 points of the winner.
- Clipping over 0.5%, or signal-to-noise under 8 dB.
- Fewer than four words: names and numbers are unreliable.
Any bubble can be re-transcribed with the higher-accuracy model from its action row, which re-runs the stored clip and replaces the scores.