Soniox STT v6

Soniox STT v6, launched 2026-11-03, is Soniox’s speech-to-text model built for voice agents. Soniox pitches it as the first STT that knows what the agent is saying and can filter the agent’s own voice.

Two linked features: Agent text input (stream the agent’s outgoing TTS text into recognition for two-sided context) and Agent voice echo cancellation (ignore agent playback while still hearing the user, including barge-in). v6 also claims higher accuracy across 60+ languages for dictation, meetings, multilingual recognition, and realtime translation, plus stronger diarization, language ID, endpointing, and context customization. Available for realtime streaming and async transcription.

Problem

Voice agents often interrupt themselves after hearing their own TTS. Classic STT only hears the user side. v6 treats agent text context and echo cancellation as first-class inputs.

Differentiation

Agent-aware STT rather than generic transcription; one model for 60+ languages across realtime/async.

Limits

• Primarily a hosted API—integrate via Soniox docs.
• Needs correct agent-text and playback wiring.
• Vendor accuracy claims should be validated on your data.

Who it’s for

Voice-agent builders, meeting products, and multilingual dictation apps fighting self-echo false interrupts.

Get started

1. Read the launch post and STT docs.
2. Stream agent outbound text into recognition and enable agent-voice echo cancellation.
3. A/B barge-in rate and entity accuracy on your calls before full cutover.