Speech recognition

Speech recognition is turning recorded or live speech into text. It is the step underneath call scoring, voice bots, voice typing and interview analysis. Its output is a written transcript of what was said, often with the time of each word and a mark of who spoke.

Here is how it works in general. A microphone records sound as a wave. The software cuts the wave into very short pieces, a few hundredths of a second each. A trained model converts these pieces into sounds and then into words, using what it learned from many hours of speech with matching text. It also uses knowledge of the language to choose between words that sound alike. The result is a transcript. Accents, background noise and mixing two languages in one sentence make the task harder, and in India many callers switch between Hindi and English in the same sentence.

Accuracy is measured by the word error rate, which counts the words that are wrong, missing or added, as a share of the words actually spoken. A system can work live, giving words as the person speaks, or later on a finished recording. It may also mark who spoke each part, which is called speaker separation. The model needs many hours of speech with matching text. Indian languages and dialects have fewer such recordings than English, so they were weaker at first. The government's Bhashini project collects data for Indian languages for this reason.

Take, for example, a customer care call. The system turns the call into text so that a manager can search it and score it; see automatic quality checks and conversation intelligence. Consider a doctor who dictates notes while seeing a patient, and the system writes them down; see AI medical scribes. Say a person calls a voice bot to ask about a delivery. The bot first turns the speech into text, then decides what to answer; see voice bots.

Related pages include speech-to-text models, accent conversion, real-time voice and speech-based IVR. It is often built on a foundation model.

How is it different from computer vision? Speech recognition works on sound and gives text. Computer vision works on pictures and gives a list of objects.