Soniox STT v6 Review: The $0.12 Speech API That Hears Past Your Agent

Voice agents have a habit of interrupting themselves. The agent speaks, its own voice leaks back through a speakerphone or laptop microphone, the speech recognizer hears it as user speech, and the agent answers its own question. Soniox STT v6, released on November 3, 2026, is built around a simple fix for that loop: let the recognizer read what the agent is saying while it listens. The model takes the agent's outgoing text as input, uses it to make sense of short replies, and filters the agent's own voice out of the transcript. Streaming costs about $0.12 an hour and async transcription about $0.10, with diarization, language identification and translation included across more than 60 languages. The open Pipecat benchmark has since added a v6 row, and the numbers are strong. The one hard date is December 15, 2026, when the v5 models retire.

What Soniox STT v6 is

Soniox STT v6 is the latest generation of Soniox's speech-to-text API, available as two models: stt-rt-v6 for live streaming over WebSocket and stt-async-v6 for recorded files over REST. Both cover more than 60 languages through one unified model, so languages can be mixed in a single conversation without switching models. That matters for products that serve callers who switch languages mid-sentence, which Soniox lists as an improved area alongside accents and code-switching.

Two limits are worth stating early. Soniox STT v6 is a proprietary hosted model that runs only through Soniox's cloud API. Soniox publishes no parameter count, and there is no self-hosted or on-device option disclosed. If your product needs to run transcription on your own servers, this is not the tool for that job.

How agent text input works

A normal speech recognizer in a voice agent hears only one side of the conversation. It gets the caller's audio and never sees the question the caller is answering. Short replies suffer most, because a word or two on its own gives the model little to work with. Soniox's own docs give three examples: "River Roast" becomes "River Road", "a tow" becomes "a tour", and a dish like "tom yum" is hard to get right from sound alone. With agent text input, the model knows the agent just offered "Maple Cafe and River Roast" and can pick the right one.

Setup is one flag in the start request, "enable_agent_text_input": true, and it is available only on stt-rt-v6, not the async model. While the agent speaks, you stream its words on the same WebSocket as the caller's audio, in small JSON frames: {"type": "agent_text", "text": "...", "end_of_segment": true}. The agent's text never appears in the returned transcript. You only get the user's speech back.

Timing is the part that needs care. The messages carry no timestamps, so the model places each piece of text where it arrives in the audio stream. Soniox's docs say to send each piece about 500 milliseconds to 1 second before the user hears it, paced by TTS timestamps rather than by the language model output, which writes a reply much faster than it is spoken. With Soniox Text-to-Speech, the return_timestamps option gives the character timestamps to pace against. If the user interrupts, flush the text that was already played, send end_of_segment: true, and drop the rest of the reply, since the user never heard it.

Agent text is billed per token at the same rate as other input text tokens, and only text that reaches the model is billed. Two guardrails are easy to miss. First, send only what the agent says out loud: strip anything the text-to-speech engine does not speak, such as SSML, markdown, or expression tags like [laughs]. Second, use agent text only for the agent's spoken words. Background information belongs in the separate context feature, which covers names and vocabulary for the whole conversation while agent text covers what the agent just said.

Echo cancellation that keeps listening

Acoustic echo cancellation is never perfect, and parts of the agent's voice leak into the microphone, especially when it starts talking. Because v6 knows what the agent is saying, it can tell that leftover echo apart from the user and leave it out. The important detail is that v6 keeps listening to the user while the agent speaks, so a caller can still interrupt. Soniox is explicit that this does not replace acoustic echo cancellation. Agent text input handles what your existing echo cancellation leaves behind, so keep it switched on.

This is also where Soniox's headline claim needs a qualifier. The company calls v6 the world's first speech recognition model that knows what your voice agent is saying. AssemblyAI shipped agent context in Universal-3.5 Pro Realtime in July 2026, and it also feeds the agent's replies into streaming transcription. The sharper difference is the echo filtering: Soniox's docs describe filtering leaked agent audio using that text, while AssemblyAI's context docs describe no such filtering. "First with model-level echo filtering" is the defensible version of the claim. "First to feed the agent's words in" is not.

What else changed in v6

Voice typing gets the most detailed treatment in the announcement. Soniox says it invested in real-world dictation data across applications, specialized vocabulary, speaking styles and languages, and the result is more accurate rendering of numbers, currencies, units, dates, alphanumeric identifiers, names and specialized terminology. Spoken commands such as "period" or "new line" are written as words, so the application can apply them. A single wrong digit changes an amount, so this is the area where a small accuracy gain is worth the most.

Meeting transcription and speaker diarization are the second focus. Soniox says training data covering rapid speaker changes, interruptions, overlapping voices, heated discussions, similar-sounding speakers and varied microphone conditions substantially improves both transcription and diarization. The real-time model covers live meetings and the async model covers completed recordings, and both return a speaker-aware record for downstream summaries and action items.

Translation is the third improvement. V6 delivers more accurate real-time translation across the same 60-plus languages and adjusts how much speech it gathers before translating, which matters in German and Korean, where late-arriving information can change how earlier words should be translated. Translation is included in the hourly rate. One formatting change needs testing before any upgrade: v6 writes "15 dollars and 20 cents" as spoken, not "$15.20", and hyphens in codes appear only when spoken.

Pricing, worked out

Soniox prices its speech-to-text API by token, and its pricing page gives a per-hour equivalent for each mode. Streaming runs about $0.12 an hour and async file transcription about $0.10 an hour. The token rates behind those figures are published on the same page: streaming charges $2.00 per million input audio tokens, $4.00 per million input text tokens and $4.00 per million output text tokens, while async charges $1.50, $3.50 and $3.50 per million for the same three token types. Soniox's usage reference says one hour of audio is about 30,000 input audio tokens and one hour of speech is about 15,000 output text tokens. The arithmetic checks out: streaming one hour costs 30,000 times $2.00 per million ($0.06) plus 15,000 times $4.00 per million ($0.06), which is $0.12. Async works out to $0.045 plus $0.0525, or $0.0975, which Soniox rounds to about $0.10.

Agent text input adds a small line of its own. Input text tokens on streaming cost $4.00 per million, so an agent speaking at roughly 150 words a minute, or about 12,000 tokens in a full hour of talking, adds about $0.05 an hour at most. In a real two-way call the agent talks for only part of the hour, so the practical figure is closer to $0.02 or $0.03. The cost is small either way, but it is not zero, and it is easy to forget when budgeting because the headline rate is per hour of audio.

Most voice agents also need a voice, and Soniox sells that too. Its text-to-speech API charges $4.00 per million input text tokens and $21.50 per million output audio tokens, or about $0.70 an hour of generated speech. A full Soniox voice stack therefore runs about $0.82 an hour of agent talk time: $0.12 for listening plus $0.70 for speaking. At November 15, 2026 market rates (1 USD = 0.7561 GBP, 1.4258 CAD), that is about £0.62 or CA$1.17 an hour, with streaming alone at about £0.09 or CA$0.17 and async at about £0.08 or CA$0.14. All conversions in this article use those November 15 rates and are approximations. There is no free tier: the pricing page lists pay-as-you-go rates with no free allowance, and AnchorTerminal's November 13 review notes new accounts have had no free credits since October 27, 2025. Rate limits are published too, according to the same review: real-time allows 100 requests a minute and 10 concurrent streams, while async allows 10 GB of storage and 1,000 files.

How it compares

The table below combines MarkTechPost's November 3 comparison of streaming prices with the live results table from the open-source Pipecat STT benchmark, which scores services on semantic word error rate and time to final segment across 1,000 real-world samples. Prices are the vendors' published rates as reported by MarkTechPost. Benchmark figures are from the Pipecat table as it stands now, judged by Claude Sonnet 5.5, and the benchmark's own note says word error rate differences under about 0.25 percentage points are not meaningful on this data.

Two readings of that table matter. First, on price, Soniox undercuts every rival in it by at least three times: AssemblyAI at $0.45, Deepgram at $0.46 for mono audio, and ElevenLabs at $0.39 are all in a different band. Soniox's own pricing page adds the big-cloud comparison: Google's Gemini 3.5 transcribe live at about $0.54 an hour and Azure Speech at about $1.00 an hour for real time, with OpenAI's batch transcription starting at $0.36 an hour and no native streaming. Those are Soniox's own figures, and they point the same way as the table.

Second, on accuracy, the v6 row is the strongest in the table and it is new. When MarkTechPost covered the launch on November 3, it reported that no v6 benchmark numbers existed yet and used the v5 scores instead. The Pipecat table now lists stt-rt-v6 at 0.74% mean semantic word error rate with 88.8% perfect transcripts and a 266 millisecond median time to final segment. On the same table, v5 sits at 1.11% and 83.6%, and v4 at 1.04% and 83.2%. The gap between v6 and v5 is 0.37 points, which clears the benchmark's own 0.25-point noise bar, so the improvement is real on this measure. The gap between v6 and Meta's muse-voice-transcribe-1.0 at 0.80% is not meaningful by the same rule, so v6 is tied at the top rather than clearly ahead of everything. One honest footnote: on this table, v5 was a small step back from v4, so v6 is the first version to beat v4.

For async work the comparison is simpler. ElevenLabs Scribe v2, the non-realtime version, runs $0.22 an hour according to AnchorTerminal's November review, or about £0.17 and CA$0.31, against Soniox async at $0.10. Deepgram's diarization add-on of about $0.0020 a minute works out to roughly $0.12 an hour on top of its base rate, which is the whole price of Soniox streaming by itself.

What is still missing

The strongest part of the v6 story is also the newest, so it deserves the most care. Soniox published no v6 accuracy or latency numbers of its own at launch. The Pipecat row supplies an independent number now, but it is one community-run benchmark on 1,000 samples, scored by a language model judge, and Soniox's own announcement still leans on the word "significantly" without a number attached. Treat the 0.74% figure as a good early signal, not a settled fact.

The pricing page is clear about rates and silent about everything else a buyer plans around: no published latency guarantee, no uptime commitment, and no volume discount table. The API is pay-as-you-go with no subscription plan, which keeps small usage cheap and leaves heavy users to negotiate. Teams that need a self-hosted deployment, an on-device model, or a published parameter count will not find any of those here. On the plus side for privacy, Soniox says customer audio and transcripts are never used for training, and real-time audio is not stored. Finally, the "world's first" wording is marketing, and the July 2026 AssemblyAI launch shows why. Soniox's real edge is narrower and more useful: it is the only service in the comparison that filters leaked agent audio at the model level using the agent's own words.

How to migrate before December 15

Soniox has deprecated the v5 models, and they will be removed on December 15, 2026. After that date, v5 requests route to v6 automatically. That sounds painless, and for the async model it mostly is, but the streaming model has a formatting change and a new feature to wire up, so the move deserves a short plan rather than a silent redirect.

First, change the model name to stt-rt-v6 or stt-async-v6. Second, if you run a voice agent, add "enable_agent_text_input": true to the streaming start request and stream the agent's text 500 milliseconds to 1 second ahead of playback, paced by your text-to-speech timestamps. Third, keep your acoustic echo cancellation switched on. Fourth, handle interruptions the way the docs describe: flush the text already played, send end_of_segment: true, and drop the rest. Fifth, test number and currency formatting against your own data, because v6 writes amounts as spoken words. Sixth, run the switch in staging first and compare transcripts side by side with v5 before you move production traffic. One more date for WebSocket users: from January 15, 2027, the API key must travel with the WebSocket connection rather than in the start message, so update the connection code in the same pass. The deadline is 30 days from the November 15 publication of this article, which is enough time for a careful switch and short enough that it should go on the sprint board now.

Pro tips and common mistakes

The most common mistake is pacing agent text by the wrong clock. Do not forward text as it comes out of the language model, and do not send the whole reply at once. Both run faster than speech. Pace by when the audio is actually played, using text-to-speech timestamps, and send small pieces often. Soniox's docs suggest sending new text after each chunk of user audio, or on a timer of about 50 milliseconds.

The second mistake is sending the wrong text. Strip SSML, markdown and expression tags before you send anything, and never use agent text for instructions or background notes. One segment covers one agent turn, not one sentence, so mark the end of the turn and not the end of each sentence. The third mistake is assuming the agent's text is free. It is billed per token at the input text rate, so a chatty agent adds a few cents an hour. Small, but worth a line in the budget for high-volume deployments.

One tip that pays off immediately is to pair agent text with the context feature rather than treating them as alternatives. Context covers the names and words of your business for the whole conversation. Agent text covers what the agent just said. Used together, they fix different classes of error: context fixes the product name the model has never heard, agent text fixes the short answer that depends on the question.

Frequently asked questions

What is Soniox STT v6?

Soniox STT v6 is the latest generation of Soniox's speech-to-text API, released on November 3, 2026, as stt-rt-v6 for live streaming and stt-async-v6 for recorded files. It covers more than 60 languages and adds agent text input, which feeds the agent's words into the recognizer, plus model-level echo filtering that removes leaked agent audio from the transcript.

How much does Soniox STT v6 cost?

Streaming costs about $0.12 an hour and async transcription about $0.10 an hour, based on Soniox's published token rates. Agent text input is billed as input text tokens at $4.00 per million on streaming, which adds up to about $0.05 an hour for a constantly talking agent and less in a normal two-way call. Diarization, language identification, translation and smart formatting are included. There is no free tier for new accounts.

Is it really the first speech model that knows what the agent says?

Not in the strict sense. AssemblyAI shipped agent context in Universal-3.5 Pro Realtime in July 2026, so feeding the agent's replies into transcription is not new. Soniox's distinct claim is the echo filtering: using the agent's text to remove leaked agent audio at the model level, while still listening for interruptions.

What happens to Soniox v5?

The v5 models are deprecated and will be removed on December 15, 2026. After that date, v5 requests route to v6 automatically. The switch itself is a model-name change, but streaming users should also test the new number formatting and wire up agent text input before the deadline rather than relying on the automatic redirect.

Can I run Soniox STT v6 on my own servers?

No. Soniox STT v6 is a proprietary hosted model available only through Soniox's cloud API, over WebSocket for streaming and REST for files. Soniox publishes no parameter count and discloses no self-hosted or on-device option. Teams that need on-premises transcription should look at self-hosted engines instead.

Does it work for meetings, or only for voice agents?

Both. The async model handles recorded meetings, and the real-time model handles live ones, with improved speaker diarization across 60-plus languages. Agent text input is the voice-agent feature, and it only works on the streaming model, but the accuracy, diarization and translation improvements apply to meeting transcription too.

Soniox STT v6 is a narrow release done well: one clear idea, a price that undercuts the field by at least three times, and a migration deadline that forces a decision. The benchmark row that appeared after launch backs up the accuracy claim, and the missing pieces are honest ones: no free tier, no self-hosting, and no vendor-published v6 numbers of its own. If you run a voice agent on Soniox v5, move before December 15 and test the number formatting first. If you run one on a rival service at three or four times the hourly rate, the next four weeks are a cheap time to find out whether the agent text trick is worth the switch.

Sources

• Soniox, "Introducing Soniox STT v6" (official announcement)
• Soniox, "Speech-to-Text and Text-to-Speech API Pricing" (official pricing page, read November 14, 2026)
• Soniox Docs, "Agent text input" (official documentation)
• MarkTechPost, "Soniox STT v6 Adds Agent Text Input and Echo Filtering for Voice Agents," November 3, 2026
• Pipecat STT Benchmark, results table (open-source benchmark, read November 14, 2026)
• AI Daily Post, "Soniox STT v6: Text Input and Echo Filtering," November 4, 2026
• AnchorTerminal, "Soniox Speech-to-Text review (2026)" (free-credit cutoff and rate limits, re-scored November 13, 2026)
• RateMorph, USD to GBP exchange rate, November 15, 2026
• RateMorph, USD to CAD exchange rate, November 15, 2026

Related reading on AI Core Updates: Mistral's voice agent model in our Voxtral 3 review, ElevenLabs' meeting teammate in our ElevenAI explainer, and the wider field in 7 AI Agents You Can Text in 2026.