Soniox releases stt v6, introducing agent text input and echo filtering for voice AI

Soniox has launched Soniox STT v6, the newest iteration of its speech-to-text API, which processes audio from voice agents. This new version, available as stt-rt-v6 for streaming and stt-async-v6 for files across over 60 languages, allows the model to process the agent's own speech. This capability enables the system to filter out audio leakage, such as echoes from speakerphones, which previously caused transcription errors. Enabling agent text input requires setting the flag 'enable_agent_text_input': true in the start request, and it is exclusive to the real-time stt-rt-v6 model. When active, the model uses the streamed agent text to separate its own voice from the caller's speech. A formatting note for users upgrading is that v6 renders spoken currency as text (e.g., '15 dollars and 20 cents') rather than numerical symbols. While Soniox has not published v6 benchmarks, the company notes that its v5 model achieved a 1.34% semantic WER on the Pipecat STT benchmark, with a median TTFS of 260 ms. Soniox states that AssemblyAI shipped agent context in Universal-3.5 Pro Realtime in July 2026, but Soniox emphasizes that its v6 specifically offers filtering for leaked agent audio. The older v5 models are slated for deprecation on December 15, 2026.