Soniox's speech-to-text update listens to both sides of a call

single source· 1 articles · confidence: low · first seen 2026-11-03 20:18 UTC

What this means for you

If you run a phone agent, echo and one-word answers are the two things that break the conversation. Soniox is fixing both at the recognition layer, but published no accuracy numbers, so test it on your own call recordings before moving traffic. Otherwise, nothing to act on.

Soniox released version 6 of its speech-to-text API on 3 November, adding an input that voice-agent builders have had to work around: the agent's own words. The model takes the agent's script or output as text alongside the caller's audio, and uses it to catch short replies that audio alone tends to miss and to strip the agent's own voice out of what it transcribes. It ships as stt-rt-v6 for streaming and stt-async-v6 for asynchronous jobs. No pricing, accuracy figures or evaluation date were published.

Key facts

  • ·Soniox released Soniox STT v6, a speech-to-text API, on 3 November 2026. source
  • ·The model accepts the voice agent's own text as an input while it processes the caller's audio. source
  • ·It ships as stt-rt-v6 for streaming and stt-async-v6 for asynchronous transcription. source
  • ·Soniox says the agent text is used to recognise short replies and to drop the agent's own voice from the audio. source
  • ·The release carried no pricing, no accuracy figures and no evaluation date. source

What the sources say

  • MarkTechPost — Press-release write-up listing the agent-text input, echo filtering and the two model names; no pricing or test results.

Sources

The original reporting. Follow these — they did the work.

← the wire