Gradium TTS streams audio in 155 ms median on Coval benchmark
Overview
Baseten reports that Gradium's text-to-speech model reaches a median Time-to-First-Audio of 155 ms, with a 2 ms interquartile range, on the Coval TTS benchmark, captured May 4, 2026.
The figures come from Baseten's first-party blog post.
Gradium's Delayed Streams Modeling design runs audio a few 80 ms steps behind incoming text, so speech can begin before the LLM has finished its response. Gradium also uses kernel fusion and CUDA Graphs, which it says cut median TTFA by more than 2.5x at small batch sizes.
Written by AI from the articles below · updated Oct 13, 10:21 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- Baseten BlogOfficialGradium TTS starts streaming audio in 155 ms median on Coval benchmark
AIGradium's text-to-speech model reaches a median Time-to-First-Audio of 155 ms with a 2 ms interquartile range on the Coval TTS benchmark, captured May 4, 2026. Its Delayed Streams Modeling design runs audio a few 80 ms steps behind incoming text, so speech can start before the LLM finishes its response. Gradium cuts latency with kernel fusion and CUDA Graphs, which Gradium says reduced median TTFA by more than 2.5x at small batch sizes.
Heat trend
Not enough continuous observations to show a trend yet.