Nari Qwen3-TTS Preview Ranked #1 in Latency on Speko

Speko, a leading platform for voice AI infrastructure and benchmarking, independently measured an experimental Nari Qwen3-TTS endpoint at 51 ms p50 and 52 ms p90 time to first audio from US East. It was the fastest endpoint in the benchmark snapshot.

Speko Text-to-Speech benchmark showing Nari Qwen3-TTS at 51 ms p50 time to first audio and approximately 1526 naturalness Elo — Speko TTS benchmark snapshot as of Aug 30, 2026.

Results

p50 TTFA: 51 ms, versus 78 ms for the next-fastest result
p90 TTFA: 52 ms, versus 85 ms for the next-fastest result
Naturalness: ~1,526 Elo, above the 1,500 field average and ahead of xAI Grok (1,495) and Rime (1,429). We also outperformed Alibaba’s vanilla Qwen3-TTS endpoint (1,310) by 216 points.

Our results are hidden on Speko’s live benchmark as it is still not generally available. Book a call for Early Access to our inference platform.