Soniox STT v6: voice agents stop hearing themselves

Soniox STT v6 fixes the bug every voice agent builder has hit: the agent hears its own voice and interrupts itself. Soniox shipped the model on November 3 with two connected features, agent text input and agent voice echo cancellation. It also set a hard date. The v5 models go away December 15, and v6 writes some numbers differently. If your voice agent parses prices or order IDs, that second part matters more than the first.

What actually happened

Per Soniox's launch post and models changelog:

• Agent text input. You stream what your agent is saying on the same WebSocket as the caller's audio (enable_agent_text_input). The model then hears both sides of the conversation, so short answers like "Faro" or "aisle, please" make sense.
• Agent voice echo cancellation. With text input on, v6 tells the agent's playback apart from the caller and drops it from the transcript, while still catching the caller when they interrupt. Soniox says to keep your normal acoustic echo cancellation on; this handles what it leaves behind.
• Both features are on the real-time model, stt-rt-v6, in every supported language. Async is stt-async-v6.
• Soniox claims better accuracy across 60+ languages, plus better diarization and real-time translation. We found no independent benchmarks yet.
• Migration: v6 is API-compatible with v5; swap the model name. After December 15, v5 requests auto-route to v6.

The catch is in the changelog's formatting notes. v6 writes "15 dollars and 20 cents" as spoken, not as $15.20. Hyphens in codes appear only when spoken, so you won't get "NP-38-BX" unless the caller said the dashes.

Why it matters for your business

Self-interruption is the top reason phone agents feel broken on speakerphone and in the car. If v6 holds up on your audio, that is a real conversion fix for booking and ordering lines. But "auto-routes with no API changes" is not the same as "no behavior changes." Any regex that expects $ amounts or hyphenated SKUs downstream of the transcript can quietly fail on December 15. Pin stt-rt-v6 now, rerun your recorded test calls, and fix the parsers on your schedule instead of Soniox's.

Sources: Soniox, "Introducing Soniox STT v6", Soniox Docs, Models and changelog.