New Time to First Audio benchmark ranks TTS latency: fastest 76ms, streaming median 376ms
ArtificialAnlys · x · 2026-11-17
- Artificial Analysis launches Time to First Audio, a benchmark measuring how quickly TTS models start speaking — the number that decides whether a voice agent feels responsive. It times audible speech (not first byte) daily on 12 fixed prompts across 4 input lengths, reporting a rolling 7-day mean.
- Key findings:
- Leading models range from 76 ms (SuperTone Gradium) to nearly a second (Gemini 3 Flash TTS at 960 ms).
- Streaming is the dividing line: 65 streaming models have a median of 376 ms, while 20 non-streaming models must finish generating first — median 2.0 s.
- Speechify Simba 3.5 starts in 116 ms, the highest Arena Elo of any model under 200 ms; Inworld Realtime TTS-2 follows at 186 ms.
- Turbo variants earn the name: ElevenLabs v4 Turbo hits 258 ms vs 653 ms for v4 at the same Arena Elo; Fish Audio Drama 3 matches at 258 ms, Cartesia Sonic 3.6 at 287 ms.
- Eleven v4 Turbo and Simba 3.5 sit on the Pareto frontier of quality vs first-audio latency.
Related event: Artificial Analysis Launches TTS Time-to-First-Audio Benchmark(4 posts)→
More from Models
- Claude's Incognito mode still reads your Gmail — economists flag misleading UX — Afinetheorem · 2026-11-17
- Benchmark finds AI rewrites overuse 'though' 5x more than humans: 0.89 vs 0.18 per 1k words — rseroter · 2026-11-17
- Blind test: Hy4 Flash beats Claude Haiku 5.5 at website generation for most humans — gnukeith · 2026-11-17
- DeepSeek V4.1 Flash on Vera Rubin NVL72: 63-93x per-chip throughput vs GB300 in vLLM benchmarks — vllm_project · 2026-11-17
- First DeepSeek V4.1 Flash results on Vera Rubin NVL72: up to 93.3x per-chip throughput vs GB300 — vllm_project · 2026-11-17
- Dev verdict on Google's new model: easier to work with than OpenAI or Anthropic, but odd mistakes — tokenbender · 2026-11-17