Speech Recognition (ASR)

Automatic speech recognition, or ASR, converts spoken audio into text. It estimates what was said from the recorded signal. It differs from text-to-speech, which produces spoken audio, and from speech translation, which expresses speech in another language.

A transcript contains decisions

Suppose a customer says, “Do not cancel order fifteen.” A transcript that drops “not” or changes “fifteen” to “fifty” can reverse the intended action. Readable punctuation does not prove that these details were captured correctly.

Recognition systems use acoustic information and learned language patterns. Radford and colleagues' speech-recognition research provides an example of learning from large audio-transcript collections. Systems vary in their architectures and supported languages; one benchmark does not establish quality for every accent or recording condition.

Measure errors that matter to the task

Word error rate compares substitutions, deletions, and insertions with a reference transcript. For a four-word reference, one substitution and no other edits gives 1 / 4 = 25%. This is an illustrative calculation. The metric weights word edits, not their business consequences, and conventions for segmentation and normalization affect comparisons.

Speaker diarization identifies who spoke when and is distinct from recognizing the words. Timestamps, speaker labels, and translation may require additional components. Preserve access to the permitted source audio and review uncertain names, numbers, and negations before using a transcript to trigger an action.

Reference: Radford et al.: Robust Speech Recognition.