# What is Speech to Text?

Also called Automatic Speech Recognition, ASR, Transcription.

Speech to text is the conversion of spoken audio into written text by an automatic speech recognition model. Systems output words along with timings and often a confidence score, and may operate on a completed recording or on a live stream. Accuracy varies with audio quality, accent, vocabulary, background noise, and how many people are speaking.

Modern recognizers process audio in short frames, learn an acoustic representation, and decode it into text with a model that also knows which word sequences are plausible. Streaming systems emit partial results that get revised as more audio arrives, which is why live captions rewrite themselves mid sentence. Batch systems see the whole recording at once and are usually more accurate.

Quality is commonly reported as word error rate, the proportion of words inserted, deleted, or substituted against a reference transcript. A single headline figure hides a lot. Error rates rise sharply for accented speech, specialist vocabulary, overlapping speakers, and telephone audio, so a benchmark number from clean read speech rarely predicts performance on real calls.

Recognition is often bundled with neighboring tasks that are technically separate: diarization, which labels who spoke when, punctuation and casing restoration, language identification, and custom vocabulary boosting for names and product terms. Confidence scores are useful as a signal to ask for confirmation, but they are model estimates and should not be read as calibrated probabilities.

For interactive use, the tradeoff between speed and accuracy is explicit. A streaming recognizer returns something usable immediately and improves it, which suits conversation, while a batch pass over the same audio produces the better transcript for the record. Many systems run both, using the live result to respond and the batch result to store.

## Key points

- Outputs text with timings and per word confidence estimates
- Streaming results are provisional and get revised
- Word error rate rises with noise, accents, and jargon
- Speaker labeling and punctuation are separate add on tasks
- Live and batch passes can serve different purposes

## In practice

A recruiter records a thirty minute screening call. The recognizer produces a timestamped transcript, tags the two speakers separately, and restores punctuation. The candidate's surname is transcribed three different ways until the vocabulary list for the role is supplied, after which it stays consistent. The recruiter then searches the transcript for 'notice period' and jumps straight to the eleven minute mark.

## Related terms

- [Voice Agent](/en/glossary/voice-agent)
- [Text to Speech](/en/glossary/text-to-speech)
- [Meeting Transcription](/en/glossary/meeting-transcription)
- [Voice Latency](/en/glossary/voice-latency)
- [Telephony Integration](/en/glossary/telephony-integration)

[Back to the AI Glossary](/en/glossary)