Glossary
What is speech recognition (ASR)
In short
Speech recognition, or automatic speech recognition (ASR), is software that turns spoken words into text: from a phone call, a voice message or a meeting recording.
AI basics · speech recognition
MMOOZZGG.com editorsChecked: 1 min read
For a child
When grandma reads out a pie recipe on the phone, you quickly write every word in your notebook. A computer can also listen and write. It catches the sounds, guesses which words hide behind them and types them on the screen. It finds this hard when the room is noisy or someone talks very fast, but it tries, and every year it gets confused less often.
In plain words
A microphone records sound as a wave. The program cuts it into short pieces, turns them into a picture of frequencies and uses it to guess the sounds, then the words. To choose between similar options, it looks at the meaning of the phrase: “buy milk” is more likely than “buy mulk”. This is how video subtitles, voice typing on a phone and transcripts of voice messages in a chat app work. Errors most often come from noise, accents, rare surnames and product names.
For business
Call transcripts open up data that used to be lost: what customers ask, where a manager did not name the price, why a deal fell through. Companies use them for quality control in a contact center, automatic meeting minutes, voice agents and search across a call archive. The cost depends on the hours of audio, the mode (real time or after the call) and where the model runs: in the cloud or on your own server. A call recording holds personal data, so it matters where it is processed and whether the customer was told about the recording. Test quality on your own calls: phone audio, industry terms and interruptions lower accuracy. The connection to your phone system is usually tested in a pilot.
For a specialist
Automatic speech recognition turns an audio signal into a sequence of words. The classic design had an acoustic model, a pronunciation dictionary and a language model. Today end-to-end models dominate: CTC, encoder-decoder models with attention, and transformers over a spectrogram. wav2vec 2.0 (Baevski and co-authors, 2020) learns speech representations from unlabeled audio. After pretraining on 53,000 hours and fine-tuning on only ten minutes of labeled data, it reaches a 4.8% word error rate on a clean test set. Whisper (Radford and co-authors, 2022) was trained on 680,000 hours of multilingual data with weak supervision and works on standard tests without fine-tuning. The main metric is word error rate (WER), the share of wrong words. Related tasks are speaker diarization, which labels each line by speaker, and language identification.
Example
A clinic transcribes all incoming calls to the front desk. Once a week the manager sees which questions are asked most often and how many calls ended without a booking.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map