Nari Labs open-sources Qwen3-TTS 1.7B: 50ms latency, 10x cheaper than ElevenLabs
alexcovo_eth · x · 2026-09-12
- Nari Labs launched what it calls the world's fastest and cheapest speech generation engine, built on the open-source Qwen3-TTS 1.7B model.
- Time-to-first-audio (TTFA) is just 50ms — 5x faster than Cartesia.
- Cost is $5 per million characters, roughly 10x cheaper than ElevenLabs.
- It reportedly beats the market average in quality and expressiveness on Speko AI benchmarks; a free trial is available now.
Related event: Nari Labs Open-Sources Qwen3-TTS 1.7B: 50ms Latency at a Tenth the Cost(2 posts)→
More from Multimodal
- AI Mourns Humanity in Dark-Comedy Short 'We Leave the Lights On', Made with Suno v6 — AIandDesign · 2026-09-12
- Turning a 3D white model into a commercial: GPT-6 Astra + Seedance 2.5 + CapCut workflow — HeyAmit_ · 2026-09-12
- Paper-Tearing Comparison Video Shows a Year of Video-Gen Physics Gains — sabage27 · 2026-09-12
- Removing "AI slop" hallmarks from images with a single prompt to Astra — floguo · 2026-09-12
- Short film 'The Fly' made in 4 hours on one RTX 5090 with ComfyUI, Minimax, Krea 2 and Suno — aurelm · 2026-09-12
- ComfyUI-QwenASR v1.1.0: official Qwen3-ASR models, smart ITN, and long-form forced alignment — Narrow-Particular202 · 2026-09-12