The post explains how passing agent text into a speech model lets it filter out the agent's own voice, useful for anyone building voice agents.
Soniox just realeased their latest speech-to-text model for voice agents, dictation, and meeting tools.
But there's something really cool about it:
Developers can send the model the audio they want to transcribe, along with the text of what the agent said to prompt that response.
Pretty clever idea.
The model can use that context to better recognize the user's reply and filter out the agent's own voice if the microphone picks it up.
Introducing Soniox STT v6, the world’s first speech recognition model that knows what your voice agent is saying and automatically filters out the agent’s own voice. Two connected innovations make this possible: → Agent text input gives recognition context from both sides of the conversation, improving its understanding of user responses. → Agent voice echo cancellation filters out the agent’s playback while continuing to listen to the user, even when they interrupt. The result: fewer false interruptions, fewer clarification loops, and more reliable conversations. V6 also delivers significantly higher accuracy across 60+ languages for voice typing, meeting transcription, multilingual recognition, and real-time translation, with stronger speaker diarization, language identification, endpoint detection, and context customization. One unified speech model. All 60+ languages. The full feature set, with languages freely mixed. Available now for real-time streaming and asynchronous transcription. Explore what’s new in Soniox STT v6: https://soniox.com/blog/soniox-stt-v6
Santiago · x.comPublished · added here
Original title: Soniox just realeased their latest speech-to-text model for voice agents, dictation, and meeting tools.
