What is speech-to-text?
Speech-to-text (also called transcription or speech recognition) is software that turns spoken words into written text, either live as someone speaks or from a recording.
It is the first step in every AI voice agent: before the AI can understand a caller, it has to hear them. It also powers call transcripts, meeting notes, captions and voice search.
What speech-to-text means (and why it matters)
Every call, meeting and voice note holds useful information that is lost unless someone writes it down. Speech-to-text captures it automatically, so you can search calls, summarise meetings, coach sales reps and let AI act on what customers say.
How speech-to-text works
| Step | What happens |
|---|---|
| 1. Capture | Audio comes from a phone line, microphone or file |
| 2. Process | The model turns the sound into likely words |
| 3. Context | It uses surrounding words to pick the right spelling ("their" or "there") |
| 4. Output | Text, often with timestamps and who said what (diarization) |
Real-time ("streaming") transcription sends text back as the person speaks, which voice agents need. Batch transcription processes a finished recording and is usually cheaper.
Costs
Prices vary by provider and mode. For example, Telnyx lists speech-to-text options from $0.0015 to $0.015 a minute on its pricing page. At those rates, transcribing 1,000 minutes of calls costs a few dollars.
How to get accurate transcripts, step by step
1. Improve the audio
Use good microphones and reduce background noise. On calls, a clear line helps most.
2. Pick the right mode
Streaming for live conversations and voice agents; batch for recordings.
3. Add custom vocabulary
Give the system your product names, staff names and industry terms.
4. Turn on speaker labels
Diarization separates the caller from the agent, which makes transcripts far easier to use.
5. Check a sample
Review a few transcripts each week and fix recurring errors.
A worked example
A sales team records every call. Speech-to-text transcribes each one, an AI summarises the key points and next steps into the CRM, and the manager searches transcripts for common objections to improve training.
Common mistakes
- Ignoring consent. Tell callers when calls are recorded or transcribed, and follow local recording laws.
- Expecting perfection on noisy audio. Fix the audio first.
- Skipping custom vocabulary. Brand and product names are the most common errors.
FAQ
How accurate is speech-to-text?
Very accurate on clear audio in common languages, and less so with heavy noise, crosstalk or rare terms. Test it on your own calls.
Is speech-to-text the same as voice recognition?
Not quite. Speech-to-text writes down what was said. Voice recognition usually means identifying who is speaking.
Can it handle different languages?
Many providers support dozens of languages, and some detect the language automatically.
Final verdict
Speech-to-text is cheap, fast and the foundation of voice AI. Get the audio right, add your vocabulary, and it turns every conversation into searchable, usable data.
Want calls answered and transcribed automatically?
We build AI voice agents that answer, transcribe and log every call into your CRM.