Soniox STT v6 Adds Agent Text Input and Echo Filtering for Voice Agents

Soniox just dropped its STT v6, and the buzz is that it finally “gets” what your voice agent is saying while it hears the caller. For anyone running a contact‑center stack or a voice‑assistant service, that sounds like a sweet deal – but the devil’s in the details. The new model ships as stt-rt-v6 for streaming and stt-async-v6 for file‑based transcription, covering more than 60 languages. The big selling point is “agent text input” and an echo‑filter that claims to strip out the agent’s own voice from the transcript. Let’s break down what that actually means for independent hosting providers, why the pricing looks tempting, and where the hidden risks lie.

What the “agent text input” actually does

The core idea is simple: you turn on a flag – "enable_agent_text_input": true – in the start request, and then you stream the exact words your TTS engine is spitting out over the same WebSocket that carries the caller’s audio. Each snippet is a tiny JSON packet like {"type":"agent_text","text":"…","end_of_segment":true}. Soniox advises you to send each piece about half a second to a second before the audio actually plays, using the TTS timestamps as a guide.

This timing trick is crucial because the model has no timestamps of its own. It simply drops the incoming text into the audio stream wherever it arrives. When a caller interrupts (the classic “barge‑in”), you flush any pending text, send an end_of_segment flag, and drop the rest. The agent’s text never shows up in the final transcript – it’s just a context cue for the model to better understand short replies like “yes” or “no”. In theory, that should solve the classic “River Roast” → “River Road” mis‑recognition that Soniox’s docs flag as a problem.

Echo filtering: hype or real value?

Echo has been a perennial headache for voice agents. Speakerphones, laptop mics, and even cheap headsets leak the agent’s own audio back into the microphone, and the STT engine mistakenly thinks the agent is the caller. Soniox claims its v6 model uses the streamed agent text to separate that leftover audio and filter it out, on top of your existing acoustic echo cancellation (AEC). In practice, you still need AEC – the model only cleans up what AEC misses.

That distinction matters. If your AEC is already solid, the incremental gain may be modest. If you’re running on cheap hardware or in noisy environments, the extra layer could shave off a few percent of false triggers. But remember, Soniox hasn’t published any latency or accuracy numbers for v6 yet, so you can’t quantify the benefit until you run your own tests.

Pricing: a tempting headline, but watch the token costs

Soniox lists a flat $0.12 per hour for streaming and $0.10 per hour for async transcription. Diarization is bundled, which is a nice perk compared to some rivals that charge extra per minute. However, the “agent text input” feature is billed as token usage – $4.00 per million input text tokens on real‑time streams. If you’re feeding a high‑volume TTS engine, those tokens can add up quickly.

For a typical contact‑center handling a few hundred calls a day, the streaming fee is a drop in the bucket. But for a large‑scale voice‑assistant platform pushing millions of TTS utterances, the token charge could become a non‑trivial line item. Compare that to AssemblyAI, which also offers agent context but at a higher streaming price and without the echo‑filter claim. The bottom line: the headline price looks cheap, but you need to model token usage before signing up.

Benchmark reality: v5 numbers and the missing v6 data

The only hard numbers we have are for v5, measured on the open‑source Pipecat benchmark. Soniox’s v5 scored a 1.34% semantic word‑error‑rate (WER), with 83.3% of transcripts being perfect, and a median time‑to‑final‑segment (TTFS) of 260 ms. Its P95 latency sat at 305 ms. Azure’s offering posted a lower mean WER at 1.21% but a much higher median latency of over a second.

Soniox hasn’t released any v6 figures, so you’re essentially buying a promise that the model is “significantly more accurate” without proof. The only way to validate the claim is to run your own Pipecat‑style test set or a real‑world pilot. Until that data lands, treat the “best price” claim as marketing spin rather than a hard advantage.

Competitive landscape: who else is doing this?

AssemblyAI rolled out a similar agent‑context feature in July 2026 with its Universal‑3.5 Pro Realtime model. It also streams the agent’s replies into the STT pipeline, but its docs don’t mention any echo‑filtering. Deepgram’s Nova‑3 models, ElevenLabs’ Scribe v2, and other players like AssemblyAI and Azure all support streaming across dozens of languages, but their pricing is higher – ranging from $0.45 to $0.55 per hour for streaming, with extra diarization fees.

The differentiator Soniox touts is the echo‑filter layer. If that works as advertised, it could be a genuine edge for low‑cost deployments that can’t afford premium AEC hardware. But the market is crowded, and the real test will be reliability at scale. A single glitch in echo filtering could cause the agent to hear its own voice and loop back, creating a nasty feedback loop that hurts user experience and raises compliance concerns.

Migration headaches and deprecation timeline

Soniox says the upgrade is a drop‑in swap: just change the model name from stt-rt-v5 to stt-rt-v6 (or the async equivalents). The older v5 models are slated for retirement on December 15, 2026, after which any v5 request will automatically route to v6. That gives you a narrow window to test the new model before the old one disappears.

However, there’s a formatting quirk you need to watch. v6 outputs “15 dollars and 20 cents” as spoken words rather than the symbol‑based “$15.20”. Hyphens in code strings only appear when spoken. If your downstream parser expects the old format, you’ll need to adjust your post‑processing logic. That’s a small change, but it’s the kind of hidden cost that can bite when you’re moving at scale.

Actionable advice for independent providers

First, run a controlled pilot. Set up a test call flow, enable enable_agent_text_input, and compare v5 versus v6 on your own data. Pay attention to edge cases – heavy accents, code‑switching, and rapid barge‑ins – because those are where echo filtering and agent‑text context will show their teeth.

Second, model token usage. Estimate the average number of words your TTS engine emits per call, multiply by daily call volume, and calculate the expected token cost. If the token fee pushes you beyond your margin, consider a hybrid approach: use v5 for low‑risk calls and only flip to v6 where short‑answer accuracy is mission‑critical.

Third, keep your AEC pipeline solid. Soniox’s echo filter is a supplement, not a replacement. A weak AEC will still let leakage through, and the model’s filtering may not catch everything, especially in noisy environments.

Finally, watch the deprecation clock. With the v5 shutdown looming in mid‑December, you have less than two months to decide. If you can’t afford the token cost or the integration effort, you might look at AssemblyAI’s context feature as a fallback, even though it lacks echo filtering. The market will continue to push down prices, but the real value lies in reliability and predictability – the two things that keep a hosting business afloat when hyperscalers start charging “premium” for every extra feature.

Bottom line: Soniox STT v6 offers a clever way to give the model a heads‑up on what the agent is saying, and it adds an echo‑filter layer that could save you a few percent of transcription errors. The price looks attractive, but the token fee and the lack of published accuracy data mean you need to test before you commit. In a world where VC‑backed hype often outpaces real‑world performance, a pragmatic, data‑driven pilot is the only way to know if v6 lives up to the buzz.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Reporting is based on the source material cited below. Sources: MarkTechPost; marktechpost.com; Global1.News (04 November 2026).

By Allan Ali, Global1.News