Nari Labs serves fast, low-cost Qwen3-TTS
Nari Labs offers optimized versions of the open Qwen3-TTS and Qwen3-ASR speech models through a simple API. The company says they lead Coval's voice AI benchmark on the balance of quality, latency and cost.
What it does
Nari Labs serves optimized versions of the open Qwen3-TTS 1.7B (speech) and Qwen3-ASR 1.7B (transcription) models through simple production APIs, with Fast and Standard endpoints.
Results and limits
On Coval's benchmark (1-day view, September 14, 2026), Nari's Qwen3-TTS Fast had the lowest word error rate (3.8%) and was #2 on time-to-first-audio at a median of 63 ms. Nari says the official Qwen3 TTS Flash Realtime endpoint scored 8.8% WER and 692 ms on the same benchmark.
The rankings exclude dedicated inference endpoints.
Worth knowing
TTS Fast costs $10 per 1M characters, and the Standard endpoint $5 per 1M. The free public beta ended September 16, 2026; use the GA model IDs.
Try it yourself
Send a test request with [a 2-sentence voiceover script] to the Fast endpoint and to your current TTS provider, then compare time-to-first-audio and how each pronounces [a tricky name].