Editorial illustration for Soniox STT v6 adds text input, echo filtering; no v6 accuracy or latency
Soniox STT v6: Text Input & Echo Filtering
Soniox pushed out STT v6 this week, the newest version of its speech-to-text API built for voice agents. The company ships it in two flavors: stt-rt-v6 for live streaming and stt-async-v6 for processing recorded files, covering more than 60 languages. The headline additions are agent text input and echo filtering, both aimed at a problem anyone who has built a phone bot already knows about.
Cascaded voice agents typically run speech recognition on the caller's audio alone. The model never sees what question the agent just asked, so short replies get mangled. Soniox's own documentation shows "River Roast" coming out as "River Road," and "a tow" turning into "a tour." Layered on top of that is echo: agent audio bleeding back through a speakerphone or laptop mic, getting transcribed as if the caller said it, and triggering the agent to interrupt itself mid-sentence.
Soniox says v6 fixes both by feeding the agent's own text into the model while it listens, letting it use that context to catch short answers and filter out its own voice. Pricing lands at $0.12 an hour for streaming and $0.10 for async, with diarization and language ID included.
Soniox has released Soniox STT v6, the latest generation of its speech-to-text API. Soniox STT v6 is a speech-to-text model that reads what a voice agent says while hearing the caller. It uses that text to recognize short replies and to drop the agent’s own voice.
Why this matters
For anyone building voice agents, the agent-text-input trick is the real story here, not the version bump. Feeding the agent's own script into the STT model so it can filter echo and parse short replies is a sensible fix for a problem every phone-bot builder hits: the system hearing itself talk. That's worth testing against your own call flows.
But we'd hold off on the hype. Soniox shipped stt-rt-v6 and stt-async-v6 across 60+ languages without a single accuracy or latency number attached. The company's own benchmarks page still cites v5 performance on the Pipecat STT benchmark, 1.34% semantic WER, 83.3% perfect transcripts, leaving v6 buyers to take the upgrade on faith.
For developers evaluating speech-to-text vendors, that's a gap worth flagging before you swap API versions in production. Echo filtering and text-input context are useful engineering, but until Soniox publishes v6 numbers on the same open benchmark, there's no way to confirm it didn't trade accuracy for those new features. Ask for the data before you migrate.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Track AI cost per active developer and merged pull request, not tokens
From Archives The Letter Company Releases Letterpress-1, a Coding Agent for Growth Engineering
Common Questions Answered
What are the two deployment options available in Soniox STT v6?
Soniox STT v6 is available in two flavors: stt-rt-v6 for live streaming audio and stt-async-v6 for processing recorded files. Both versions support more than 60 languages and are designed specifically for voice agent applications.
How does the agent text input feature in Soniox STT v6 solve the echo filtering problem?
The agent text input feature allows the STT model to read what the voice agent says while hearing the caller's audio, enabling it to recognize short replies and filter out the agent's own voice from the recording. This solves the common problem where phone bots hear themselves talking, which every voice agent builder encounters.
What is the main limitation of Soniox STT v6's release according to the article?
Soniox released stt-rt-v6 and stt-async-v6 across 60+ languages without publishing any accuracy or latency metrics. This lack of performance data means users should test the new version against their own call flows before fully adopting it.
Why is the agent-text-input feature considered more significant than the version bump itself?
The agent-text-input feature addresses a fundamental problem that all phone-bot builders face: the system hearing its own voice during conversations. Rather than just being a version increment, this feature provides a practical solution that improves the quality and reliability of voice agent interactions.
Further Reading
- Soniox STT v6 Adds Agent Text Input and Echo Filtering for Voice Agents - MarkTechPost
- Introducing Soniox STT v6 - Soniox Blog
- Speech-to-Text & Speech Recognition API - Soniox
- Speech-to-text API for AI voice agents - Soniox
