Nari Labs

Nari Labs is opening up a streaming Speech-to-Text endpoint aimed at developers building real-time voice applications like agents, meeting tools, and live captioning. The service runs Qwen3-ASR 1.7B, an open-source ASR model developed by the Qwen team at Alibaba Cloud that supports multilingual speech, music, and song recognition along with language detection and timestamp prediction , served on the company's own inference engine. The pitch on this launch is a 40 ms time-to-final-segment and pricing of $0.06 per hour of audio, positioned as the fastest and cheapest streaming transcription endpoint available, with a Top 3 placement on the Coval accuracy benchmarks. The launch matters because streaming ASR latency directly caps how natural a voice agent can feel, and current hosted options for Qwen3-ASR sit much higher on the latency curve. For reference, Baseten's own Qwen3-ASR-1.7B streaming deployment reports P95 finalization latency under 0.50 seconds across up to 64 concurrent streams on RTX 6000 and H100 GPUs , so a 40 ms TTFS claim on the same underlying model is a meaningful step for teams stitching together listen-think-speak loops. Nari Labs is a two-person startup that introduced Dia, a 1.6 billion parameter text-to-speech model designed to produce naturalistic dialogue directly from text prompts , and has since been working down the latency stack across the voice pipeline. The team, led by Toby Kim, previously shipped Nari Qwen3-TTS, a high-performance single-H100 serving implementation of Qwen3-TTS 1.7B CustomVoice that exposes streaming and non-streaming speech generation over HTTP plus WebSocket-based input streaming, targeting sub-50 ms time-to-first-audio at 10 requests per second . This ASR endpoint is the listening half of that same effort, giving builders a matched-latency pair for full-duplex voice systems.