Voice Agents' Self-Echo Issue: Soniox STT v6 Leverages Agent Output as a Filtering Signal
Soniox has released STT v6, adding text input support and echo filtering capabilities for voice agents. With context awareness of what the agent is currently saying, the model can clearly pick up short user responses like "uh-huh" or "okay" while suppressing its own voice. Both streaming and asynchronous versions are launched at the same time.
Voice agents have a long-standing awkward pain point: they need to keep listening while they are speaking. Traditional STT systems often accidentally pick up the agent's own output from the speaker, which messes up the recognition result when the caller only sends a short 1-word response like "okay" or "mm".
Soniox announced STT v6 on November 3rd, introducing a new dedicated input to solve this problem: the exact text that the voice agent is outputting. As stated in the official release, STT v6 reads the transcript of the agent's ongoing speech while listening to the caller. It uses this reference text to identify short user responses and discard the agent's own voice from the recognition output. The logic behind this design is straightforward: the model gets an extra context signal that tells it exactly which words it just spoke, and which words come from the end user.
There are two official release variants: stt-rt-v6 for real-time streaming use cases, and stt-async-v6 for asynchronous workflows. Developers building real-time voice agents will prioritize the streaming version, while teams working on audio recording, post-processing or offline transcription will use the asynchronous one. The official announcement has not published metrics including latency, accuracy benchmarks, and pricing, so it is too early to draw any conclusions. Details on how to pass the agent text through the API, toggle the echo filter on/off, and its performance across different acoustic environments are all subject to official documentation and real-world testing.
For end users, this type of improvement delivers very tangible benefits: voice assistants will less frequently repeat themselves, cut off users mid-speech, or ignore short responses such as "sure", "go ahead" and "wait a second". For developers, it adds a usable new context signal, and shifts echo processing from a pure audio post-processing step one layer deeper into the model's input pipeline.
The original release post is short, and we have covered all the key information above. The full article link is attached here: Soniox STT v6 Adds Agent Text Input and Echo Filtering for Voice Agents
发布时间: 2026-11-04 04:18