HYPE OR FACT? · CHECKED 3 NOVEMBER 2026
Soniox STT v6: the "world's first" speech model that knows what your voice agent says and filters out its voice?
OUR VERDICT
Partly factOur original verdict: “Fact or hype: partly fact. The features are real and documented: Soniox STT v6 takes your voice agent's spoken text alongside the caller's audio, uses it to understand replies, and filters leftover echo of the agent's voice out of the transcript. But "world's first" is overstated. AssemblyAI's streaming model has taken the agent's text as context since June 2026, and Speechmatics' voice SDK could already ignore an assistant's enrolled voice. Doing both from the agent's text in one model may be new, but we can't confirm a first. "Significantly higher accuracy" comes with no v6 numbers yet. We haven't tested it.”
THE CLAIM
“Introducing Soniox STT v6, the world's first speech recognition model that knows what your voice agent is saying and automatically filters out the agent's own voice. ... V6 also delivers significantly higher accuracy across 60+ languages.”
What’s actually true
- What's new: Soniox, a speech AI company led by founder and CEO Klemen Simonic, released STT v6 on 3 Nov 2026 as stt-rt-v6 (real-time) and stt-async-v6 (files). The headline feature is "agent text input": while your voice agent speaks, you stream its words to Soniox on the same WebSocket as the caller's audio (enable_agent_text_input).
- It does two things, per Soniox's docs. First, the model knows what the caller is answering, so short replies are recognised better: Soniox's examples are "River Roast" not "River Road", and "a tow" not "a tour". Second, because it knows what the agent is saying, it can tell leftover echo of the agent's voice apart from the caller and leave it out of the transcript, while still hearing the caller if they interrupt.
- It isn't a replacement for echo cancellation. Soniox's own docs say: "Keep your echo cancellation on. Agent text input handles what it leaves behind; it doesn't replace it." Standard acoustic echo cancellation, like the kind built into WebRTC calls, still does the main job.
- Why "world's first" is overstated: AssemblyAI's changelog shows Universal-3.5 Pro Realtime launching on 23 Jun 2026 as a streaming model that "takes agent context as input". Its docs describe an agent_context parameter for passing the voice agent's spoken replies so the model "knows both sides of the dialog". Speechmatics' open-source voice-agent examples (December 2025) show its voice SDK ignoring an enrolled assistant voice labelled __ASSISTANT__ to stop the agent's own speech being transcribed. That SDK is now deprecated.
- What may be new: Soniox's version streams the agent's words in time with the audio and uses that text, not a voiceprint, to filter the echo, inside the speech model itself. We found no identical feature from another provider, but that's not enough to confirm a world first.
- "Significantly higher accuracy" has no numbers yet. The v6 blog and changelog give none. Soniox's benchmarks page (dated August 2026) still shows the older v5 model on the open-source Pipecat STT benchmark (1,000 voice-agent clips). There Soniox v5 had 1.34% mean semantic word error rate, behind Azure at 1.21% but ahead of most others. These are figures Soniox publishes; we haven't rerun them.
- "60+ languages": Soniox's supported-languages page lists 60 languages, so it is 60 rather than more than 60. Soniox says agent text input works in every supported language.
- Price: about $0.12 per hour of real-time transcription and $0.10 per hour for files, billed in tokens, with diarization, language identification and translation included. Agent text is billed as input text tokens ($4.00 per 1M on real-time).
- Practical advice for small businesses building phone bots or voice agents: this is aimed at the problem you actually hit, which is bots that mishear short answers or interrupt themselves on speakerphones. Plumbing matters: Soniox says to send each piece of agent text about 0.5 to 1 second before the caller hears it, using your TTS word or character timestamps, not when the LLM writes it. Test v6 on your own call recordings before switching.
- Watch the automatic upgrade: v5 is retired on 15 Dec 2026 and requests then go to v6 with no code change. v6 writes amounts as spoken ("15 dollars and 20 cents" instead of "$15.20") and drops unspoken hyphens in codes, which could break code that reads prices or reference numbers. We haven't tested v6 and haven't watched Soniox's 64-second launch video.
Sources
- Soniox blog: Introducing Soniox STT v6 (3 Nov 2026)
- Soniox docs: agent text input
- Soniox docs: models and changelog (v6, v5 retirement, formatting changes)
- Soniox docs: supported languages (60 listed)
- Soniox pricing
- Soniox benchmarks (v5 on the Pipecat STT benchmark, August 2026)
- AssemblyAI docs: Conversation Context (agent_context)
- AssemblyAI changelog: Universal-3.5 Pro Realtime "takes agent context as input" (23 Jun 2026)
- Speechmatics Academy: voice agent speaker ID, ignoring the __ASSISTANT__ voice
- The claim: Soniox (@soniox_ai) on X (3 Nov 2026)
Cost
About $0.12 per hour real-time and $0.10 per hour for files (token-based), with diarization, language ID and translation included. Agent text adds input text tokens at $4.00 per 1M on real-time. Text-to-speech is about $0.70 per hour.
USE IT IF
Developers and small teams building voice agents, phone bots, dictation or multilingual meeting tools who want cheap, fast speech-to-text and can wire up TTS timestamps.
SKIP IT IF
Anyone who needs independent proof of v6 accuracy before switching (none is published yet), or who uses an all-in-one voice platform that doesn't let you choose the speech-to-text provider.
How we knowChecked on 3 November 2026. Verdicts: Fact = the claim holds up; Partly fact = real, but with caveats or overstated parts; Hype = the headline claim isn't supported. Products change quickly, so check the maker's site before you buy. Our methods & sources ↗