Soniox STT v6 is a speech-to-text release for developers building voice agents. On the real-time model, stt-rt-v6, teams stream the text an agent is about to speak into recognition on the same WebSocket as the caller's audio. The model transcribes each reply knowing what the caller is answering, which helps most with short and ambiguous answers, and it filters the agent's playback out of the transcript while still listening if the caller interrupts. Voice agents have usually run recognition on caller audio alone. When speakerphone, laptop, or in-car playback reaches the microphone, the model can treat the agent's voice as user speech, which leads to false interruptions and feedback loops. Agent voice echo cancellation handles leftover agent audio after acoustic echo cancellation, and Soniox says to keep that processing on. Both agent features are limited to stt-rt-v6 and work in every language Soniox supports. The async model, stt-async-v6, is for file transcription. Soniox, founded in 2020 and based in Foster City, offers speech-to-text, text-to-speech, and translation across 60+ languages. Founder and CEO Klemen Simonic, whose work includes Stanford, Google, and Meta, leads the company with co-founder and chief architect Ambroz Bizjak. v6 also updates voice typing, meeting transcription, multilingual recognition, and real-time translation, including speaker diarization, language identification, and endpoint detection. The v5 models are deprecated and set for removal on December 15, 2026, after which those requests route to v6.