ASR Confidence Scores and Calibration

Speech recognizers produce words and internal scores for candidate outputs. A product may expose confidence for a word, phrase or entire utterance to indicate uncertainty. The score can be based on acoustic evidence, alternative hypotheses, model probabilities or a separate confidence estimator. Research on word confidence and calibration shows that these approaches have different behavior. A high raw score does not by itself mean “90 percent chance the word is correct.” That interpretation requires testing against labeled audio from the relevant setting. Calibration asks whether predictions at a given confidence level are correct about that often. For example, if many words labeled 0.8 are right only half the time, the scores are overconfident. A reliability plot and suitable calibration metrics can expose this. Calibration may drift when microphones, accents, noise or vocabulary change. Even a calibrated average can hide poor results for rare names or numbers. A threshold chosen for general captions may be too permissive for a medication amount or a payment command. Confidence is useful for allocating attention. An editor can prioritize uncertain words, a voice interface can ask a clarifying question, and a data pipeline can flag spans for review. It should not be used as a substitute for the source audio in consequential cases. The score may reflect that a fluent word sequence is common rather than that every syllable was heard. An error detector also makes mistakes: high-confidence wrong words can pass, and correct unusual names can be flagged. Evaluate a confidence system on independent audio, including both errors it catches and correct words it wrongly rejects. Report calibration by relevant group or condition, and test the full review workflow rather than one numeric score. Keep the recording or a compliant audit trail when a person must verify a disputed transcript. An honest interface explains what confidence is based on and allows correction.