Artificial Analysis Launches Time to First Audio Benchmark for TTS Models

Artificial Analysis announced a Time to First Audio benchmark measuring how quickly text-to-speech models begin speaking, with leading models ranging from 76 ms to nearly a second. The benchmark tests every model daily on 12 prompts across four input lengths, timing audible speech and reporting a rolling 7-day mean. Streaming models showed a median of 376 ms versus 2.0 seconds for non-streaming models.
Event version 1.
Key insights
Simba 3.5 leads under-200ms tier
SpeechifyAI's Simba 3.5 starts speaking in 116 ms, the highest Arena Elo of any model beginning under 200 ms. InworldAI's Realtime TTS-2 follows at 186 ms.
Turbo variants cut latency at same quality
ElevenLabs' Eleven v4 Turbo starts in 258 ms versus 653 ms for Eleven v4 at the same Arena Elo. Fish Audio Drama 3 matches at 258 ms and Cartesia Sonic 3.6 is at 287 ms.
Streaming divides latency performance
The 65 streaming models begin speaking within a few hundred milliseconds, a median of 376 ms. The 20 non-streaming models must finish generating before returning anything, a median of 2.0 seconds.
Benchmark methodology
Every model is measured daily on the same 12 prompts across four input lengths, timing audible speech rather than the first byte and reporting a rolling 7-day mean.