Speech-to-Text
Speech-to-text, also called automatic speech recognition, is the technology that converts spoken audio into written text so that software and people can search, analyse and act on what was said.
Definition
Speech-to-text takes an audio signal and produces a transcript. Modern systems use neural networks trained on large amounts of recorded speech, and they predict the most likely words given both the sounds and the context. The output can include timestamps, punctuation and, in some tools, a label for each speaker.
Accuracy is usually expressed as a word error rate: the share of words that are wrong, missing or added compared with a human-made reference transcript. It varies a lot with audio quality, background noise, accents, speaking speed, specialist vocabulary and the language itself. Real-time transcription also has to balance speed against accuracy.
Transcripts feed everything that comes next in a voice system, so a recognition error can change the meaning of an answer, especially for names, numbers and technical terms. Good practice is to test on your own audio and accents, keep the ability to listen to the original recording, and not treat a transcript as a perfect record.
Key points
- It converts speech into text; it is also called automatic speech recognition (ASR).
- Accuracy is measured by word error rate against a human transcript.
- Noise, accents, specialist words and fast speech reduce accuracy.
- Names and numbers are the usual weak points, so check the audio for anything important.
- It is the first step of most voice AI agents, before language understanding.
Example
A training centre records ten-minute check-in conversations with learners. The transcripts are searchable, so a coordinator can find every conversation that mentions a transport problem. Names of towns are misspelt in a few transcripts, so the team checks the audio before acting on them.
Frequently asked questions
Is speech-to-text the same as voice recognition?
Not exactly. Speech-to-text works out what words were said. Voice or speaker recognition works out who is speaking. The two can be combined but they are different tasks.
How accurate is speech-to-text?
It depends on the audio and the language. Clear recordings in widely spoken languages give good results, while noise, strong accents or technical jargon reduce them, so test on your own recordings.
Is a transcript personal data?
If it can identify a person, yes. Transcripts of conversations should be protected, kept only as long as needed and handled under the data protection rules that apply to you.
Related terms
AI conversations, human decisions
Passisto holds structured voice conversations with candidates, employees, clients and students, and turns them into reports your team can act on.