Baseten adds Gradium TTS, a real-time voice model with sub-human first-audio latency
Baseten said in a blog post that Gradium TTS, a text-to-speech model built for real-time voice agents, is now available in its Model Library as a dedicated deployment. The model uses Delayed Streams Modeling, a framework from Kyutai that aligns text and audio to a shared 80 ms clock and runs the audio stream a few decoding steps behind the text, so it can start speaking before an upstream LLM has finished its response. Baseten cites Coval benchmark data captured May 4, 2026 showing a median time-to-first-audio of 155 ms with a 2 ms interquartile range, and says kernel fusion, CUDA Graphs and its own weight-delivery network cut cold starts by 2–3x.
For voice-agent builders, first-audio latency below typical human response times gives models room to reason or retrieve context without breaking conversational flow.