Glossary
Speech-to-text (STT)
Speech-to-text is the automatic conversion of spoken language into text.
The technology is also called automatic speech recognition (ASR). In a phone assistant it is the first step: only once it is clear what was said can the system reply. On the phone this is harder than with a dictation feature, because the audio is narrowband and background noise, dialects and poor connections come on top.
Errors occur especially with proper names, streets, email addresses and sequences of digits. Good systems therefore have important details confirmed or spelled out. Besides real-time recognition during the call there is transcription after the call, from which logs and summaries are produced. What counts for quality is therefore not only the error rate of a model but the result in your own use. Transcripts contain personal data and fall under data protection law.
In practice
You notice recognition quality less in single words than in what follows from them: is the name in the summary correct, was the right number noted, did the assistant understand the question? It can be improved in three places. The assistant repeats important details and has them confirmed. Technical terms, product names and frequent surnames from your business are part of the stored knowledge, so the system can place them. And for delicate details such as email addresses there is a second route, for example clarifying them during the callback.
An example
At a veterinary practice a pet owner says: “This is Mrs Szymański, it’s about Balu, he hasn’t eaten since yesterday.” Recognition first turns the surname into something quite different. Because the assistant asks her to spell the name and repeats the callback number digit by digit, both are correct in the note. Without that check the practice would not have found the dog’s record and might have dialled a wrong number.
A common misunderstanding
An accuracy figure from a data sheet is frequently mistaken for one’s own result. Such figures come from clean recordings, not from a call made in a car or on a building site. Just as common is the assumption that a transcript is a verbatim record. It remains an automatic transcription with errors. For anything binding, such as appointments, addresses and amounts, what counts is the detail confirmed in the call and not the raw text.
At Call0
At Call0, calls produce transcripts and summaries. The privacy policy describes how they are processed.
Read on:
On this topic:
As of: 12 October 2026