Soniox STT v6: Agent text input and echo filtering

What changed

Soniox released STT v6, a speech recognition API built for voice agents that introduces agent text input and agent voice echo cancellation. The company deployed the model in two variants: stt-rt-v6 for real-time streaming and stt-async-v6 for batch file processing, supporting 60 or more languages.

Agent text input allows developers to stream the agent’s outgoing text into the recognition engine while the caller speaks. A single flag enables this feature in the start request, and agent words are sent as JSON frames on the same WebSocket connection that carries caller audio. The model uses this text as context to recognize short caller replies more accurately, such as names, numbers, and domain-specific terms that carry ambiguity when heard in isolation.

Agent voice echo cancellation leverages that text to separate and filter out the agent’s playback that leaks through speakerphones, laptops, or other devices, preventing the agent from responding to its own words. The model continues listening to the caller during agent playback, so callers can interrupt without triggering false recognition events.

Beyond voice agents, v6 provides higher accuracy for voice typing and meeting transcription. The company improved recognition of numbers, currencies, units, dates, alphanumeric identifiers, and names. It strengthened speaker diarization, language identification, endpoint detection, and handling of accents, code-switching, and language switching within sentences. v6 has not yet published benchmark results.

Why it matters

For voice agents, these two capabilities address practical challenges. Short caller replies like “Faro,” “aisle,” or “8” are inherently difficult to distinguish without context. A voice agent that asks “Porto or Faro?” and then hears only “Faro” lacks information to confirm which option was selected. Agent text input gives the STT model both sides of the conversation, reducing misrecognition on those critical short responses.

Echo presents a familiar operational problem. When an agent’s voice leaks back into the microphone, the STT model treats it as caller speech and may transcribe it. The agent then reacts to its own words, causing self-interruption or feedback loops. While acoustic echo cancellation (AEC) reduces this problem, v6 uses semantic information to filter out what acoustic processing misses.

For voice typing and meeting transcription, stronger recognition of specialized vocabulary and improved speaker diarization reduce manual corrections and increase reliability of action items. Better multilingual handling also matters for global products, though the source provides no concrete v6 benchmarks to support any specific accuracy claim.

Migration is straightforward for existing users. Soniox deprecated v5 and will retire it on December 15, 2026, after which v5 requests will route to v6 automatically. Users can adopt v6 by changing the model name in their requests.

What to test

Before adopting v6, test several behaviors and validate key claims:

Accuracy gains on your data: Soniox states v6 is “significantly” more accurate than v5 but has published no v6 benchmarks. Its v5 predecessor achieved 1.34 percent semantic word error rate on the open-source Pipecat benchmark, behind v4’s 1.25 percent. Run v6 on your own voice agent calls and transcription data to measure whether the improvement justifies migration effort.

Agent text input timing: Messages carry no timestamps, so the model places text where it arrives in the audio stream. Soniox recommends sending each piece 500 milliseconds to 1 second before playback, synchronized with TTS timestamps. This window is critical; send too early and context may not align with caller speech, too late and the model has already passed that moment in the audio. Test this in your TTS workflow to find what works.

Echo cancellation performance: v6 does not replace acoustic echo cancellation; it handles residual agent audio that existing AEC leaves behind. Keep AEC enabled. Test how much leftover agent audio v6 filters out on your specific hardware, whether speakerphones, laptops, or in-car systems.

Formatting behavior: v6 writes “15 dollars and 20 cents” as spoken, not “$15.20”. Hyphens in codes appear only when spoken. If your application depends on normalized number or code formatting, you may need post-processing logic.

Barge-in handling: On caller interruption, you must flush unplayed agent text and send end_of_segment: true. Verify this works smoothly in your call flow without dropping genuine caller speech.

Context effectiveness: The model accepts context such as product names, professional vocabulary, and project terminology. Test whether supplying context actually improves accuracy for your domain, since the benefit varies by use case.

Comparison with AssemblyAI: AssemblyAI’s Universal-3.5 Pro Realtime also ships agent context as of July 2026. Compare latency, accuracy, and pricing between the two. Soniox does not publish v6 latency numbers, making direct comparison more difficult. The clearer difference is echo filtering: Soniox describes using agent text to filter leaked agent audio, while AssemblyAI’s documentation describes no such filtering.

The conclusion

Soniox STT v6 addresses a genuine problem in voice agent design. Models typically hear only the caller and struggle with short replies that lack context. Agent text input and echo cancellation are practical solutions to this constraint. Both features are billed separately and require explicit flag changes to enable, so adoption is an active choice rather than a breaking change.

The main limitation is the absence of published v6 benchmarks. Soniox’s v5 scored 1.34 percent semantic word error rate on Pipecat, behind Azure’s 1.21 percent but within a competitive range. The accuracy claim for v6 remains unverified until independent results appear.

Pricing at $0.12 per hour streaming and $0.10 per hour async is competitive. Diarization and language identification are included, and agent text input is billed at $4 per 1 million tokens, a low rate relative to its value. Deployment requires only a model name change, and the December 15, 2026 deprecation of v5 gives users time to plan.

Watch for independent benchmarking of v6 on Pipecat or similar datasets. Also track whether voice agents using agent text input and echo cancellation report fewer clarification loops and self-interruptions in production. Real-world validation matters more than vendor claims when adoption affects user experience.

For teams building voice agents at scale, the features merit testing against your own call data. For voice typing and meeting transcription, the improvements to specialized vocabulary and speaker diarization are incremental; test them, but do not expect dramatic accuracy shifts without running your own benchmarks against your data.

Related reading

• AssemblyAI-TTS-1: $15 per 1M characters for voice agents
• Genie 4 adds real-time audio and physics to world models
• Letterpress-1: Coding agent for growth engineering