Soniox has updated its speech-to-text API to version 6, introducing features designed to let voice agents hear themselves while processing user input and filtering out their own audio.
The update ships as stt-rt-v6 for streaming calls and stt-async-v6 for batch files. It covers more than 60 languages. Soniox hosts the model and does not offer a self-hosted option.
How agent text input works
Normally, a speech-to-text engine only hears the person on the other end of the line. It misses the question the agent is currently asking. This causes errors with short phrases. In Soniox’s documentation, “River Roast” transcribes to “River Road”, and “a tow” becomes “a tour”.
The new version fixes this by accepting text input alongside audio. You enable it with a single flag in the start request: "enable_agent_text_input": true. This feature is only available on the streaming model.
While the agent speaks, you stream its words on the same WebSocket connection as the caller’s audio. Each message is a small JSON frame containing the text and a flag indicating if the segment has ended.
Timing is critical. The model places text where it arrives in the audio stream, so messages carry no timestamps. Soniox recommends sending each piece 500 ms to 1 second before playback, paced by the text-to-speech timestamps.
During barge-in, where a user interrupts the agent, you flush what was already played, send the end-of-segment flag, and drop the rest. The agent’s text never appears in the final transcript returned to the application.
Agent voice echo cancellation
With agent text input active, the model knows exactly what the agent is saying. It uses this information to separate leftover agent audio from caller speech and remove it.
The system keeps listening during playback, so callers can still interrupt. This feature does not replace standard acoustic echo cancellation. It handles what existing echo cancellation leaves behind, so you must keep standard AEC enabled.
Other changes in v6
- Better voice typing for numbers, currencies, units, dates, IDs, names, and specialized terms.
- Spoken commands like “period” or “new line” are written as words, allowing your app to apply them.
- Improved meeting transcription and speaker diarization for both live and recorded sessions.
- Stronger handling of accents, code-switching, and language switching within a single sentence.
One formatting change requires testing before upgrading. Version 6 writes “15 dollars and 20 cents” as spoken text rather than “$15.20”. Hyphens in codes appear only when spoken.
How does Soniox compare on benchmarks?
Soniox has not published accuracy or latency numbers for v6. Its benchmarks page still reports v5 on the open-source Pipecat STT benchmark. This test uses 1,000 real-world samples and measures semantic word error rate plus time to final segment.
On this metric, stt-rt-v5 scores 1.34% semantic word error rate, 83.3% perfect transcripts and 260 ms median time to final segment. Its P95 latency is 305 ms. Azure posts the lowest word error rate mean at 1.21%, but at 1,016 ms median latency.
| Feature | Soniox stt-rt-v6 | AssemblyAI Universal-3.5 Pro Realtime | Deepgram Nova-3 | ElevenLabs Scribe v2 Realtime |
|---|---|---|---|---|
| Model size | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
| Streaming languages | 60+ | 18 | 60+ (Nova models) | 90+ |
| Streaming price / hour | $0.12 | $0.45 | $0.46 mono, $0.55 multi | $0.39 |
| Agent-side text context | Yes, timed agent_text | Yes, agent_context | Not disclosed | Not disclosed |
| Model-level agent echo filtering | Yes | Not disclosed | Not disclosed | Not disclosed |
| Diarization cost | Included | Not disclosed | +$0.0020/min | Not disclosed |
| Pipecat semantic WER mean | 1.34% (v5; v6 untested) | 1.44% | 1.71% (nova-3-general) | 3.16% |
| Pipecat perfect transcripts | 83.3% (v5) | 84.7% | 76.5% | 81.3% |
| Pipecat TTFS median | 260 ms (v5) | 282 ms | 247 ms | 281 ms |
Pipecat figures come from the Pipecat STT benchmark README. Soniox numbers are for v5, since v6 has not been benchmarked yet.
Is Soniox really the first?
Soniox calls v6 the first model that knows what your voice agent is saying. That claim needs a qualifier.
AssemblyAI shipped agent context in Universal-3.5 Pro Realtime in July 2026. It also feeds the agent’s replies into streaming speech-to-text.
The clearer difference is echo. Soniox’s docs describe filtering leaked agent audio using that text. AssemblyAI’s context docs describe no such filtering.
Deployment, pricing and migration
v6 is a drop-in swap: change the model name to stt-rt-v6 or stt-async-v6.
The v5 models are deprecated and will be removed on December 15, 2026. After that, v5 requests route to v6 automatically.
Pricing is token-based. Soniox lists about $0.12/hour streaming and $0.10/hour async. Agent text is billed as input text tokens, at $4.00 per 1M on real-time.
What it means
Developers building voice agents can now prevent transcription errors caused by the agent speaking over the user. This removes the need to manually inject text into the audio stream to help the model understand context.
Users of speakerphones or laptops with poor acoustics will see fewer instances of the agent’s own voice being transcribed as user speech. The system handles the residual echo that standard cancellation misses.
The cost remains low at $0.12 per hour for streaming, and diarization is included in the price. This makes the feature accessible without a significant budget increase.
