Show HN: Nari Qwen3-TTS and Qwen3-ASR – High accuracy, low latency and cost

Nari's Qwen3-ASR Fast model achieved a #1 ranking in Time-to-Final-Segment (TTFS) with a p50 of 44 ms. Its Word Error Rate (WER) was 3.6%, placing it #2 behind AssemblyAI’s Universal 3.5 Pro, which had a WER of 3.5%. These benchmark values were based on Coval’s 1-day view as of September 14, 2026, at 15:00 UTC, with WER pooled across datasets. The rankings exclude dedicated inference endpoints.

This report highlights Nari's Qwen3-ASR Fast model achieving the #1 ranking in Time-to-Final-Segment, unlike other models that may prioritize only Word Error Rate.