Skip to main content
Cartesia has three ways to transcribe speech. Most voice agents should start with Realtime STT (Auto), which detects the end of each turn for you.

Overview

Realtime STT (Auto)

Realtime STT (Auto) (/stt/turns/websocket) streams audio over a WebSocket and has built-in turn detection. A user turn is one stretch of user speech that your app responds to once. We call this endpoint Auto because our model decides when each user turn is over.

Realtime STT (Manual)

Realtime STT (Manual) (/stt/websocket) is for apps that already run their own voice activity detection (VAD), which tells speech apart from silence and noise. This endpoint gives you tight control over when transcripts come back. When the user stops speaking, send "finalize". The model then returns the rest of the transcript right away. We call this endpoint Manual because your app decides when each user turn is over.

Batch STT

Batch STT (/stt) transcribes pre-recorded audio in a single request. It takes the whole recording at once, while the realtime endpoints only take audio at the speed it’s spoken: one second of audio per second.

Comparison

ink-2 supports English, French, Hindi, Japanese, and Spanish. ink-preview supports 44 languages. Neither needs a language parameter. Only ink-preview reports detected languages.

How to decide

  • Building a voice agent: start with Realtime STT (Auto).
  • Your app already knows when to start and stop transcribing, or you want tight control over when transcripts come back: use Realtime STT (Manual), and send "finalize" whenever the user stops speaking.

Where to go next

Understand turn detection

See how user turn events work in voice agents

Troubleshoot model behavior

Transcription errors, high latency, server errors

Check out some code examples

Example scripts for each endpoint