Soniox STT v6 Filters Voice Bots' Own Speech

Speech recognition company Soniox released Soniox STT v6 on Tuesday, Nov. 3, a model built for AI voice agents that takes in the agent's own replies as text and uses them to strip the agent's voice out of what it hears. The company calls it the "world's first speech recognition model that knows what your voice agent is saying and automatically filters out the agent's own voice."

Key takeaways

• Soniox STT v6 ships as two models, stt-rt-v6 for real-time streaming and stt-async-v6 for recorded files, both released Nov. 3.
• New "agent text input" and "agent voice echo cancellation" let the model use what the bot is saying to understand short replies and ignore the bot's own voice.
• Listed prices stay at $0.12 an hour for real-time and $0.10 an hour for async; v5 models will be retired Dec. 15, 2026.

What is agent text input in speech-to-text?

Most voice agents chain three systems: speech-to-text listens, a language model decides what to say, and text-to-speech talks back. The listening model usually hears only the caller, with no idea what question was just asked.

With v6, developers stream the agent's outgoing text to Soniox as it goes to text-to-speech, over the same WebSocket connection as the caller's audio. Soniox's documentation names the setting enable_agent_text_input. The company says the added context helps with short, ambiguous answers such as names, numbers and a plain "yes" or "no."

How agent voice echo cancellation works

On speakerphones, laptops and in-car systems, a bot's own voice often leaks back into the microphone. That can make the agent "hear" itself, interrupt itself or loop. Soniox says v6 uses the agent's text to separate that playback from the caller's speech and filter it out of the transcript.

The model keeps listening while the agent talks, so callers can still cut in. Soniox's docs advise keeping standard acoustic echo cancellation turned on alongside the new feature rather than replacing it.

Other upgrades in Soniox STT v6

The release covers more than 60 languages in one unified model, and both agent features work in every supported language. Soniox cites better handling of numbers, currencies, dates and alphanumeric IDs in voice typing, improved speaker diarization trained on overlapping speech, and more accurate real-time translation that adjusts how much speech it buffers for languages such as German and Korean.

The company claims "significantly higher accuracy" but published no v6 benchmark figures or latency numbers in its announcement. Its benchmarks page, dated August 2026, still lists the earlier stt-rt-v5: a 1.34% mean semantic word error rate and a 260-millisecond median time to final transcript on the open-source Pipecat benchmark, where Azure posted the lowest error rate at 1.21%.

The competitor figures are Soniox's own comparison of public pay-as-you-go rates and may not match every configuration.

Context: the race to fix voice agents

Voice is becoming a main channel for AI customer service, and accuracy on short turns is where agents often fail. Company research from ElevenLabs published Nov. 2 found voice agents outperforming chat on long calls, while larger platforms are pushing cheaper speech models, including Google's new voice models that talk while they think.

Soniox has moved quickly. Its v4 models arrived in January and February 2026 and v5 in June, according to its model changelog.

What's next

Soniox says v6 is API-compatible with v5, so developers can upgrade by changing the model name. The stt-rt-v5 and stt-async-v5 models are deprecated and will be removed Dec. 15, 2026, after which requests route to v6 automatically. Independent benchmarks of v6, and of the echo filter in noisy real-world calls, have yet to appear.

FAQ

What is Soniox STT v6? It is Soniox's latest speech recognition model, released Nov. 3, 2026, in real-time (stt-rt-v6) and asynchronous (stt-async-v6) versions covering more than 60 languages.

Do developers need to rewrite code to use it? No. Soniox says v6 is API-compatible with v5; agent text input is switched on with the enable_agent_text_input setting.

How much does Soniox STT v6 cost? Soniox lists real-time transcription at $0.12 an hour and async transcription at $0.10 an hour.

References

• Soniox STT v6 — Soniox (Nov. 3, 2026)
• Models — Soniox Docs
• Pricing — Soniox
• Benchmarks — Soniox (August 2026)