The Top Speech Alternative Has Its Own Confidence Score

A voice assistant hears “send the report to Mara” and shows a transcript with confidence 0.82. That number should not make every word equally trustworthy. A mistaken name can matter more than the rest of an otherwise plausible sentence.

In Google Cloud Speech-to-Text V1, a recognition result may contain several transcript alternatives. The first is the recognizer’s top-ranked hypothesis. The result’s alternative-level confidence estimates how likely its recognized words are correct in aggregate; Google says this value is set only for the top alternative of a non-streaming result or a final streaming result. Ranking also uses signals beyond this score, so the top alternative need not have the highest confidence value. The request guide and V1 response reference describe these boundaries.

The API can also provide word-specific information. With enableWordConfidence set in the V1 recognition configuration, the top result includes words with confidence values; the default is false. This gives an application a more local signal for a name or command verb. It does not turn the alternative-level score into a guarantee about any particular word. The REST reference says confidence is not guaranteed to be accurate or always supplied. Its default 0.0 is a sentinel for an unset value, so code should not silently interpret zero as a measured failure.

Live transcription adds another field with a different job. Interim results can change as more audio arrives. Their stability estimates how likely that partial text is to change; it does not estimate correctness. The guide says stability is absent on final results, while alternative confidence is typically available only on the final top hypothesis. A stable partial phrase can still be wrong, and a low stability value says little about the eventual transcription.

For a voice action with a costly mistake, display interim words as provisional and wait for a final result before using confidence. If the action depends on a person’s name, inspect available word confidence and ask the user to confirm the name when evidence is weak or absent. Keep a separate path for missing scores. Any numerical threshold should be chosen from labeled examples of the application’s own accents, noise, names, and error costs; the documentation defines fields, not a universal safe cutoff.