Soniox shipped STT v6 on November 3. Its headline voice agent feature, agent text input, is off until you set enable_agent_text_input and stream your agent's words as agent_text messages timed to playback, and the v5 models it replaces disappear on December 15.
Soniox released STT v6 on November 3, 2026 as two new models: stt-rt-v6 for real time transcription and stt-async-v6 for asynchronous jobs. The release is aimed squarely at voice agents, and its most useful new capability does nothing at all unless you turn it on.
The feature is called agent text input. Your voice agent already knows exactly what it is about to say, because it generated the words before sending them to text to speech. With v6 you can stream those words into the same WebSocket that carries the caller's audio. Soniox uses them to transcribe short or ambiguous replies more accurately (a caller who says "the second one" makes far more sense when the model knows the agent just offered two choices) and to filter the agent's own voice out of the transcript while still tracking the caller, including when the caller interrupts.
To enable it, set enable_agent_text_input to true in the start request. It is available only on stt-rt-v6. Then send the agent's words as WebSocket text frames, alongside the caller's binary audio frames, each shaped as an agent_text message with a text field and an end_of_segment flag that defaults to false. Send text exactly as the agent will say it, spaces included, and set end_of_segment to true when the agent's turn ends or is interrupted. If you send an agent_text message without enabling the feature, Soniox returns the error "Agent text requires enable_agent_text_input in the start request."
Timing is the detail most teams will get wrong. Soniox recommends sending each piece of agent text about 500 ms to 1 second ahead of audio playback, not as the LLM generates it. Text to speech runs faster than real time, so text streamed straight from the model can arrive well ahead of the audio the caller actually hears. Follow the playback, not the LLM. And keep your acoustic echo cancellation on: Soniox says agent text input handles what echo cancellation leaves behind, and does not replace it.
There is also a deadline. Soniox says v6 works with the v5 API and upgrading only requires changing the model name, but stt-rt-v5 and stt-async-v5 will be removed on December 15, 2026, with requests routed to v6. Make the switch on your own schedule, and test agent text input while you do, rather than discovering the new behavior after the automatic cutover.
Try it today
- Change your model name from stt-rt-v5 to stt-rt-v6 (or stt-async-v5 to stt-async-v6) before December 15, 2026.
- Add "enable_agent_text_input": true to the real time start request.
- Stream the agent's words as {"type": "agent_text", "text": "..."} frames on the same WebSocket, exactly as spoken, spaces included.
- Time each frame about 500 ms to 1 second ahead of audio playback, following playback rather than LLM output.
- Send end_of_segment: true when the agent's turn ends or is interrupted, and leave acoustic echo cancellation on.