Soniox STT v6: Agent Echo Cancellation, API Pricing and Migration

Soniox released STT v6 on November 3, adding agent text input and voice echo filtering to its speech recognition API. A voice agent can now tell the transcription model what it is saying, helping it understand short replies and avoid mistaking its own playback for the caller.

The useful change is in the connection between parts of a voice system. A language model may know the conversation, while the speech recognizer hears only the caller's audio. Giving that recognizer the question being answered can prevent a mistake before it reaches the language model.

What Soniox STT v6 adds

The launch announcement introduces real-time and asynchronous models across more than 60 languages. Soniox also reports improvements in meeting transcription, speaker separation, multilingual recognition and translation. Those are company claims, rather than independently established gains for v6.

The two new agent features work together in real time. Agent text input supplies the words the assistant is speaking. Agent voice echo cancellation uses that information to filter playback that leaks into the microphone, while continuing to recognize the caller when they interrupt.

Think of a caller answering a choice between two restaurants. A short restaurant name can sound like an ordinary phrase when heard alone. The preceding question gives the recognizer a better starting point. That matters because an incorrect name can send the rest of the call down the wrong path even if the language model reasons correctly.

The integration detail that matters: playback timing

Soniox's agent text input documentation makes this more specific than sending a conversation transcript alongside audio. Developers enable enable_agent_text_input on stt-rt-v6 and send small agent_text messages on the same WebSocket as the caller's audio.

Those messages should arrive roughly 500 milliseconds to one second before the corresponding speech is played. Follow actual playback, not the speed at which the language model generates text. Sending the entire answer immediately would give the speech recognizer words the caller has not heard yet.

If the caller interrupts, send any already-played text that has not been forwarded, end the segment and discard the unplayed remainder. Existing acoustic echo cancellation should stay enabled. Soniox describes the new filtering as handling the echo that remains, not replacing the audio layer.

This adds a small but important responsibility to the application: it needs a reasonable view of what has actually been spoken. Teams with buffered audio, slow connections or interruptible playback should test that synchronization before judging the model.

Soniox STT v6 pricing

Soniox's API pricing is token-based. Its approximate equivalents are $0.12 per hour for real-time transcription and $0.10 per hour for asynchronous files. These are useful estimates, not a flat ceiling on every call.

Token type	Real-time, per million tokens	Async, per million tokens
Input audio	$2.00	$1.50
Input text	$4.00	$3.50
Output text	$4.00	$3.50

Agent text is billed at the input-text rate. Extra context and output therefore belong in the cost estimate, especially for an assistant that speaks much more than the caller. A quieter transcription workload and a long, talkative support call will not necessarily have the same effective hourly bill.

Migrating from Soniox v5

The model changelog lists stt-rt-v6 and stt-async-v6 as active. Existing applications can change the model identifier without changing the underlying API. The agent-text functionality still needs the additional integration described above.

The v5 models are scheduled for removal on December 15, 2026, after which requests using their names will automatically route to v6. That makes testing worthwhile even for teams that do not plan to adopt the new agent features immediately.

There is also a formatting change. V6 preserves spoken currency wording rather than automatically converting it into a currency symbol and amount. Hyphens in codes appear only when spoken. Applications that extract prices, booking references or account identifiers should check their parsing assumptions before the automatic transition.

API compatibility does not guarantee identical output. A request can keep succeeding while a downstream parser becomes less reliable.

What still needs testing

Soniox's published benchmark page still reports v5 results, so those numbers should not be presented as evidence of v6 performance. The launch does not provide a quantified v6 accuracy or latency improvement that settles the comparison.

For a voice-agent team, the useful evaluation is a set of complete conversations: short answers, overlapping speech, speakerphone echo and interrupted replies. Count incorrect names and false interruptions alongside transcription errors. A cleaner transcript is valuable, but the product outcome is whether the call reaches the right result without unnecessary clarification.

The next step for voice infrastructure is better coordination between recognition, reasoning and playback. Soniox v6 gives the recognizer more context. Whether that improves a real call now depends partly on how carefully the application keeps those parts in sync.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn product updates and engineering decisions into clear stories that developers can understand and share.

Book a 30-minute call.