Nari Qwen3-TTS and Qwen3-ASR - Top Accuracy, Lowest Latency and Cost

Nari Labs' speech-service results illustrate why latency needs a precise definition. For its hosted Qwen models, the company reports a median 44 milliseconds from a transcription finalization request to final text, and 63 milliseconds from submitted text to the first generated audio. Those measure different stages of a voice application, rather than the entire wait for a spoken answer. Nari's benchmark account uses a September 14, 2026 snapshot of Coval's tests.

The reported rankings also separate speed from accuracy: the transcription endpoint placed first for latency and second for word error rate, while speech generation reversed those positions. The supported finding is a competitive combination of speed and transcription accuracy within that dated comparison.