Soniox announced STT v6 on November 3, 2026. For developers, playback coordination and transcript handling deserve attention before deployment.
Migration and parsing
The active models are stt-rt-v6 and stt-async-v6. Soniox says that after December 15, 2026, requests using stt-rt-v5 and stt-async-v5 automatically route to their respective v6 replacements. Existing IDs remain usable.
Check downstream parsing: v6 retains currency words, inserts hyphens in identifiers only when spoken, and returns spoken punctuation and formatting commands as words for the application to interpret.
Coordinate with playback
Agent text input sends spoken agent words alongside microphone audio on one WebSocket, providing recognition context and helping filter residual playback.
Agent text input and agent-voice echo filtering are real-time-only; retain acoustic echo cancellation. Send text roughly 0.5–1 second before actual playback, not when the LLM or TTS generates it. Close each turn with end_of_segment=true. On interruption, send any already-played unsent text, close the segment, and do not send the unheard remainder. Agent text is billed at the input-text token rate.
Suggested acceptance tests
These are our suggested checks, not benchmark results:
- Run the same recorded dialogue through both configurations. Log the local playback position beside each outgoing text message. Investigate drift using that timeline rather than judging success from a readable transcript alone. Agree on a reviewer and an acceptable deviation beforehand.
- Interrupt near the beginning, middle and end of a reply. Compare the audio the user heard with the text the service received, and record whether the next exchange behaves as expected. Save a failing example to rerun after a fix.
- Feed transcripts into a staging parser. Compare resulting amounts, identifiers and editing actions with manually reviewed expectations. Record failures separately from recognition errors so the migration review can identify which component needs work.