How agent STT works
Agent STT is a WebSocket interaction pattern for conversational applications. You stream audio in and receive speaker-labeled, turn-based transcription back: clean, punctuated, and ready to pass directly to a language model.
Turn detection runs on the server. You choose a profile based on your use case, and the server decides when to finalize each speaker's turn.
Agent STT is available on SaaS on Cloud for evaluation and feedback. It is not production-ready and not ready to scale. See Preview limitations.
What agent STT provides
Agent STT uses the Linden 1 model and gives you four things a raw transcript stream does not:
- Turn detection — the server decides when a speaker has finished talking, so you know when to respond
- Segmentation — transcripts arrive as complete speaker turns rather than as a word-by-word stream
- Diarization — each segment carries a speaker label, and speakers persist for the session
- Speaker focus — choose which speakers your application acts on and which it ignores
How agent STT differs from streaming
Both agent STT and streaming send live audio over a WebSocket, and both use the Realtime processing mode. They differ in what comes back and in who decides when a turn has ended.
Use streaming when you want the raw word-by-word stream and your own segmentation. Use agent STT when you want each finished turn handed to you ready for a language model.
Agent STT provides the speech-to-text layer only. It is not a conversational pipeline: it does not include a language model or speech synthesis.
Three separate things configure turn detection across Speechmatics surfaces, and they are not interchangeable:
- Agent STT profiles (
adaptive,agile,smart,external) — server-side, selected in the endpoint path - Voice SDK presets (
FAST,FIXED,ADAPTIVE,SMART_TURN,EXTERNAL,SCRIBE,CAPTIONS) — client-side, over the Realtime API - Realtime API
conversation_config— a silence threshold you set yourself
The only genuine correspondence is between the agent STT external profile and the Voice SDK EXTERNAL preset: both hand the turn boundary to your application. Otherwise, do not map one name onto another.
Where agent STT runs
Agent STT runs on SaaS on Cloud. Audio is processed in the EU or the US. You cannot select a region during Preview, and agent STT is not available in the AUS region.
For the features Linden 1 supports, see Feature availability.
Preview limitations
Agent STT is a preview offering with the following limitations.
- Do not send live production traffic. The service is less stable than production endpoints.
- There are no uptime or performance SLAs.
- There are no data residency guarantees. Audio may be processed in the EU or the US.
- Features may change. Preview features may be withdrawn, or may never be released publicly.
Give feedback
This is a preview, and your feedback shapes what goes to general availability. Tell us what works well, which features you use, whether something did not work as expected, a profile that behaved differently from what you anticipated, or a feature you would want before we ship more broadly.
Specific areas of interest:
- Integration experience, including documentation, SDKs, and API messages and metadata
- Accuracy and latency, including data capture where relevant, such as phone numbers or spelled-out names and account numbers
- Turn detection, and your experience with the different profiles
- Any missing capabilities that would make your product better
- What would stop you using this in production
To talk to us in person, contact your Speechmatics contact or use the channel shared in your preview welcome email. You can also fill in this form.
Next steps
- Quickstart — open a session and receive your first turn
- Turn detection profiles — choose a profile
- Agent STT messages — every message in a session