Introduction

Soniox just cut a new speech recognition generation aimed at voice-agent builders, not just meeting bots. On November 3, 2026, the company published Introducing Soniox STT v6 and flipped the docs changelog to stt-rt-v6 (real-time WebSocket) and stt-async-v6 (async). This is a closed hosted ASR API — no open weights — but it is runnable today if you already call Soniox, and the upgrade is a model-name swap.

The headline is two linked features: agent text input and agent voice echo cancellation. The model is meant to know what your agent just said, recognize short ambiguous replies more cleanly, and keep the agent’s own playback out of the transcript while still listening for barge-in.

What Shipped

  • Vendor / product: Soniox STT v6
  • Modality: speech-to-text (real-time + async), with real-time translation and diarization improvements
  • Open vs closed: closed API (no open weights disclosed)
  • Runnable IDs today: stt-rt-v6 and stt-async-v6 on the Soniox API; SDKs tagged around v2026.11.03
  • Agent text input: stream the agent’s outgoing text into recognition (flag: enable_agent_text_input) so v6 hears both sides of the conversation
  • Agent voice echo cancellation: with agent text input on, v6 filters the agent’s playback from the transcript while still listening to the user — including interruptions
  • Languages: both agent features run across 60+ supported languages on one unified model
  • Also improved: voice typing / dictation (numbers, currencies, units, dates, IDs), meeting transcription + speaker diarization, accents / mid-sentence language switching, real-time translation
  • Formatting note: v6 writes prices and alphanumeric codes closer to spoken form (less forced $15.20-style normalization)

Docs model page: Models | Soniox Speech AI Docs.

Drop-in Upgrade and Deprecation

v6 is API-compatible with v5. Swap the model string:

{ "model": "stt-rt-v6" }

{ "model": "stt-async-v6" }

Deprecation: stt-rt-v5 and stt-async-v5 are marked deprecated and will be removed on December 15, 2026. After that date, Soniox says v5 requests auto-route to v6 with no API change required — but you should still move your configs so you’re not depending on the alias.

Why It Matters for Makers

Most STT stacks still only hear the caller. That hurts short answers (“Faro”, “aisle”, “eight”) that only make sense after the agent’s last prompt, and it hurts speakerphone / laptop setups where TTS playback leaks into the mic. Soniox is betting that giving the recognizer the agent’s text — then using that to cancel agent voice — is the fix for false self-interruptions.

If you ship voice agents, voice typing, or multilingual meeting products on Soniox, this is the model to A/B this week before the December cutoff.

Original Source

Conclusion

STT v6 is not a new open checkpoint — it’s a hosted ASR bump with a clear maker hook: agents that know their own lines. Point real-time sessions at stt-rt-v6, turn on enable_agent_text_input, and test a barge-in on a speakerphone before December 15.

—Anabel ♡