Contents
At a glance
- Category
- Voice and messaging
- Also called
- STT, ASR, automatic speech recognition
- Role in a voice agent
- First stage - its errors cascade into every later step
- Streaming
- Required for live conversation; endpointing decides when the caller has finished
- Batch
- More accurate, right for voicemail, call review and note-taking
- Cheapest accuracy win
- Custom vocabulary with your product, street and staff names
What speech to text does
Speech to text - also called ASR, automatic speech recognition - turns audio into text. In a voice agent it runs first, and its output is what the language model reads. That ordering has an important consequence: a transcription error becomes a reasoning error, and no amount of model quality recovers a postcode that was heard wrong.
It is also useful outside voice agents. Transcribing sales calls, voicemails and site visits turns audio nobody has time to listen to into text an automation can summarise, classify and file - often a better first automation than a live agent because nothing happens in real time and mistakes are cheap.
Streaming versus batch
Two modes with different jobs. Streaming transcribes as the person speaks, emitting interim results that get revised, and it is required for live conversation because the agent needs to know when the caller has finished. Batch transcribes a complete recording afterwards, is more accurate because it can use the whole context, and is the right choice for voicemail, call review and note-taking.
The hard problem in streaming is endpointing: deciding the caller has actually stopped rather than paused to think. Cut in too early and you interrupt them; wait too long and the conversation drags. Most voice platforms expose a silence threshold, and tuning it against real calls is one of the highest-value adjustments in a voice deployment - see latency budget and barge-in.
What actually affects accuracy
Audio quality first: phone calls are narrowband and lossy, so accuracy on a call is always worse than on a clean recording, and a caller on speakerphone in a van is worse again. Then accents and dialects, where coverage varies substantially by provider. Then domain vocabulary - trade terms, product names, drug names, local place names - which is exactly what your business cares about and exactly what generic models get wrong.
Most providers support a custom vocabulary or keyword boost, and using it is the single cheapest accuracy improvement available. Add your product names, your street names, your staff names. Beyond that, look for speaker diarization if you need to know who said what, and word-level confidence scores, which let you route a low-confidence field to a human rather than acting on a guess.
Cost and error handling
STT is billed per minute of audio and is generally modest next to telephony and model costs. Streaming is usually priced higher than batch, and some providers charge for the connection time rather than the speech within it, which matters when callers pause a lot.
Design for mishearing explicitly. Read critical values back - "that is oh-seven-seven-double-oh, nine-oh-oh-one-two-three, is that right?" - because confirmation is far cheaper than a wrong booking. Where a value is safety-critical or expensive, prefer keypad entry over speech, and hand off to a person when confidence is low rather than proceeding on a guess.
What you should know
Test on your own recordings before choosing a provider. Vendor accuracy figures are measured on clean read speech, and your reality is a customer on a motorway with a regional accent asking about a product with an unusual name. Twenty real call recordings will separate the providers far more decisively than any published benchmark.
Speech to text FAQ
What is speech to text?
Software that converts spoken audio into written text, also called ASR or automatic speech recognition. It is the first stage of a voice AI agent, and it is also useful on its own for transcribing voicemails, sales calls and site visits into something automations can process.
What is the difference between streaming and batch transcription?
Streaming transcribes as the person speaks and is required for live conversation; batch transcribes a finished recording and is more accurate because it can use the full context. Use streaming for calls and batch for voicemail, review and note-taking.
Why does my voice agent mishear product names?
Because generic models were not trained on your vocabulary. Almost every provider supports a custom vocabulary or keyword boost - add your product names, staff names and local street names. It is the cheapest accuracy improvement available and usually takes an hour.
How do I handle a transcription error safely?
Read critical values back to the caller for confirmation, use keypad entry for anything expensive or safety-critical, and act on word-level confidence scores where the provider offers them by routing low-confidence fields to a person rather than proceeding on a guess.
Why does the agent interrupt callers?
Endpointing - the silence threshold that decides the caller has finished speaking - is set too aggressively. Most voice platforms expose it. Tuning it against real recorded calls, rather than against your own test calls, is one of the highest-value adjustments in a voice deployment.
How do I choose a speech-to-text provider?
Test on your own call recordings, not on published benchmarks. Vendor accuracy figures come from clean read speech; your reality is a customer on a motorway with a regional accent. Twenty real recordings will separate providers far more decisively than any leaderboard.
Primary source
Checked August 2026. Anything priced or rate-limited moves, so verify against the source before you budget.
Want this built rather than explained?
Knowing what speech to text means is one thing; having it running in your business is another. We scope one workflow at a time, quote a fixed price in writing after a free consult, build it, and hand it over with documentation you own.
