Nari Labs achieves sub-50 ms TTS latency with Qwen3-TTS

Nari Labs' Qwen3-TTS 1.7B CustomVoice implementation hits sub-50 ms p95 time-to-first-audio at 10 RPS on a single H100, costing ~$2 per 1M characters vs ElevenLabs V3's $100/1M. It's the only one of five engines to achieve sub-50 ms p95 TTFA.