What is Soniox STT v6 (stt-rt-v6 and stt-async-v6)?
Soniox STT v6 is a pair of speech-to-text models released on 3 November 2026: stt-rt-v6 for real-time streams and stt-async-v6 for recorded files. They replace the v5 models, which are removed on 15 December 2026, after which requests are routed to v6 automatically. The API is unchanged; upgrading means changing the model name.
The headline addition is for voice agents. With agent text input switched on, the developer streams the text the agent is about to speak into the recogniser, so it hears both sides of the conversation; agent voice echo cancellation then strips the agent's own playback from the transcript while still catching a caller who interrupts. Both work in every supported language on the real-time model.
Speaker labels, language identification, formatting and translation across 60+ languages are included in the hourly price. One behaviour change needs care: v6 writes prices and alphanumeric codes closer to how they are spoken.
How much does Soniox STT v6 (stt-rt-v6 and stt-async-v6) cost?
| Plan | Price | What you get |
|---|---|---|
| Real-time (stt-rt-v6) | About $0.12 an hour | $2.00 per 1M audio tokens; $4.00 per 1M text tokens |
| Async (stt-async-v6) | About $0.10 an hour | $1.50 per 1M audio tokens; $3.50 per 1M text tokens |
Real-time about $0.12 an hour ($2.00 per 1M audio tokens); async about $0.10 an hour ($1.50 per 1M audio tokens). Text tokens $4.00 or $3.50 per 1M. Diarization, language ID and translation included.
Is Soniox STT v6 (stt-rt-v6 and stt-async-v6) worth it?
Soniox's case starts with the price. At about $0.12 an hour for live audio, with speaker labels and translation included, it costs roughly a quarter of AssemblyAI's $0.45 agent model. A thousand hours a month comes to $120.
The echo filter is the clever part. A voice agent's own speech often leaks back through a caller's speakerphone and gets transcribed as if the caller said it; by feeding the agent's outgoing text to the recogniser, v6 knows what to discard without going deaf to an interruption.
The catch: accuracy is contested. On the open Pipecat benchmark, as published by Soniox, stt-rt-v6 had the lowest semantic error of 23 systems at 0.74%, but its margins over Meta (0.80%) and AssemblyAI's Universal-3.6 Pro (0.89%) sit inside the test's confidence interval, and four systems returned final text faster. AssemblyAI's own voice-agent test put v6 at 6.02% against 4.53% for its newest model. And the change to spoken-style prices and codes can quietly break parsers downstream.
Our take: at these prices v6 is the sensible default for high-volume transcription and agents that talk over speakerphones. Where a single misheard digit is costly, run your own calls through it and one rival before committing.
Where it earns its place
A restaurant-booking voice agent taking calls from car speakerphones
Stop the agent's own greeting being transcribed as the caller's words
Enable agent text input on stt-rt-v6 and send each TTS sentence with sendAgentText; echo cancellation filters it while barge-in still works. 5,000 call hours a month cost about $600.
A market-research agency running interviews in 12 languages
Transcribe and translate recordings without per-language setup
Batch files through stt-async-v6 at about $0.10 an hour, with speaker labels and translation in the same call.
Before you commit
- Setup
- Change the model name to stt-rt-v6 or stt-async-v6; for agents set enable_agent_text_input and stream the agent's text alongside the audio.
- Integrations
- Soniox REST and WebSocket APIs; JavaScript SDK (v2026.11.03 adds sendAgentText and an India region); community integrations such as Home Assistant.
- Data and privacy
- Weights are not published and Soniox does not advertise self-hosting. The pricing page does not cover retention or training use; check Soniox's terms before sending regulated audio.
- Cost at scale
- One hour is about 30,000 audio tokens, so 100,000 hours of live audio cost about $12,000; async about $10,000.
- Limits
- v5 removed 15 December 2026. Echo cancellation and agent text input are real-time only. Spoken-style output for prices and alphanumeric codes.
What’s good
- About $0.12 an hour live, $0.10 async, extras included
- Agent text input and echo cancellation for voice bots
- Lowest semantic error on the Pipecat benchmark, 0.74%
- Drop-in upgrade; v5 traffic routes to v6 after 15 December
What’s not
- Pipecat lead is inside the confidence interval
- 6.02% on AssemblyAI's agent-call test, behind STT-3.7
- Prices and codes now written as spoken
- No self-hosting; retention terms not on the pricing page
Best for
Teams transcribing large volumes of calls or meetings in many languages, and voice-agent builders whose callers use speakerphones.
Soniox STT v6 (stt-rt-v6 and stt-async-v6) alternatives
Soniox STT v6 (stt-rt-v6 and stt-async-v6): common questions
- Is Soniox STT v6 (stt-rt-v6 and stt-async-v6) worth it?
- Transcription priced like a rounding error, with one genuinely new idea for voice agents; the accuracy claims depend on whose benchmark you read. We rate it 4.2 out of 5.
- How much does Soniox STT v6 (stt-rt-v6 and stt-async-v6) cost?
- Real-time about $0.12 an hour ($2.00 per 1M audio tokens); async about $0.10 an hour ($1.50 per 1M audio tokens). Text tokens $4.00 or $3.50 per 1M. Diarization, language ID and translation included.
- What are the downsides of Soniox STT v6 (stt-rt-v6 and stt-async-v6)?
- Pipecat lead is inside the confidence interval; 6.02% on AssemblyAI's agent-call test, behind STT-3.7; prices and codes now written as spoken; no self-hosting; retention terms not on the pricing page.
- Who is Soniox STT v6 (stt-rt-v6 and stt-async-v6) best for?
- Teams transcribing large volumes of calls or meetings in many languages, and voice-agent builders whose callers use speakerphones.
- How much does Soniox STT v6 (stt-rt-v6 and stt-async-v6) cost for a team or in production?
- One hour is about 30,000 audio tokens, so 100,000 hours of live audio cost about $12,000; async about $10,000.
- Is Soniox STT v6 (stt-rt-v6 and stt-async-v6) safe for company data?
- Weights are not published and Soniox does not advertise self-hosting. The pricing page does not cover retention or training use; check Soniox's terms before sending regulated audio.
- What are the best Soniox STT v6 (stt-rt-v6 and stt-async-v6) alternatives?
- The closest alternatives we have reviewed are AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent), ElevenLabs Scribe v3 and Scribe v3 Live, Deepgram Nova-3 and Voxtral 3 (Transcribe and TTS).
Answers checked Nov 21, 2026.