Soniox STT v6 listens to what a voice agent says to filter out its own speech

Speech recognition company Soniox released version 6 of its Speech-to-Text model on November 3. Its main pitch is aimed at developers building voice agents: the model can be told what the agent is saying, and it uses that information both to understand the user better and to keep the agent's own voice out of the transcript. Soniox describes v6 as the first speech recognition model with this kind of awareness, a claim that has not been independently assessed.

How the agent-awareness feature works

According to the announcement, the feature has two linked parts. First, developers send the agent's outgoing text to the speech-to-text service at the same moment it goes to the text-to-speech engine. The model then has context from both sides of the conversation, which Soniox says helps it interpret short or ambiguous replies from the user, such as names, numbers or technical terms.

Second, v6 uses that same text to separate the agent's playback from the user's voice and drop the playback from recognition. The model keeps listening while the agent speaks, so if a user cuts in, their words are picked up while the agent's audio is filtered out. Soniox says the aim is to reduce false interruptions and feedback loops on devices where the speaker's output leaks back into the microphone, including speakerphones, laptops and in-car systems. Both functions are offered in every supported language.

Other changes and availability

Soniox says v6 is more accurate across more than 60 languages, with improvements in voice typing, meeting transcription, multilingual recognition and real-time translation, as well as speaker diarization, language identification and end-of-speech detection. For translation, the model adjusts how much speech it waits for before translating, which the company says matters particularly for German and Korean.

The model comes in two variants, one for real-time streaming and one for recorded files, and both are available now. The announcement itself contains no accuracy, error-rate or latency figures; it points to a separate benchmarks page, so the size of the claimed gains cannot be judged from the release alone.

On its pricing page, Soniox lists real-time transcription at $0.12 per hour of audio, with the calculator naming the stt-rt-v6 model, and asynchronous transcription at $0.10 per hour. The announcement does not mention any price change for v6, nor does it describe API changes or migration steps from earlier versions.