Speech-to-Text (STT)
Technology that converts spoken audio into written text, also called automatic speech recognition (ASR), used for transcription, captions and voice interfaces.
Definition
Speech-to-text systems take an audio signal and produce the words that were spoken, with timestamps. Modern systems are neural models trained on very large amounts of transcribed speech; open models such as Whisper and hosted APIs from the major cloud providers are common choices. Related steps include diarisation (who spoke when), punctuation and formatting, language detection and translation.
Two modes matter in practice. Streaming recognition returns words as they are spoken, for live captions and voice assistants, at some cost in accuracy. Batch recognition processes a complete recording and is more accurate and cheaper. Quality is measured by word error rate (WER), and the biggest influences on it are audio quality, vocabulary and the number of overlapping speakers. Transcripts are often passed to an LLM for summaries, action items or search.
Example: a conference records every session. Batch transcription runs overnight, diarisation separates panellists, and an LLM produces a summary and key quotes for each session. Attendees search the transcripts the next morning.
Why it matters for your business
- Meetings, calls and events become searchable, summarisable text instead of unused recordings.
- Accessibility: live captions and transcripts serve attendees and customers who need them.
- It is the front end of every voice agent; its errors propagate into everything downstream.
- Recordings are personal data. Consent, retention and where the audio is processed need deciding first.
How Teksmobile uses it
Our speech-to-text service covers model selection, diarisation and the LLM layer on top, and we have built session transcription into our own Eventify platform, as described in the event speech-to-text case study. For conversational interfaces it is paired with our voice AI work.
Related terms
Part of the Teksmobile AI Glossary. See also AI for events.
Frequently asked questions
How accurate is speech-to-text now?
On clear audio in major languages, modern models approach human transcription quality. Accuracy drops with background noise, crosstalk, strong accents, technical vocabulary and distant microphones. Measure word error rate on your own recordings.
What is diarisation?
Working out who spoke when, so a transcript is attributed to speakers. It is a separate step from recognition and is often the harder problem in meetings and panels.
Real-time or batch?
Live captions and voice agents need streaming recognition with low latency. Recordings can be processed in batch, which is cheaper and usually more accurate because the model sees the whole audio.