Nari Labs ships sub-50ms Qwen3 TTS + ASR endpoints, cheapest yet
Built on the open Apache-2.0 Qwen3 speech models, they top Coval's voice benchmarks — and the TTS engine is open source, so you can self-host instead.
Copy markdown
Sub-50ms voice at $5 per million characters
Nari's hosted Qwen3-TTS returns first audio in under 50ms server-side and Qwen3-ASR in under 40ms — against roughly 300ms for ElevenLabs TTS and 90ms for Deepgram STT. Standard pricing is $5 per 1M characters (TTS) and $0.06 per hour (ASR); the Fast tiers run $10 and $0.12.
Top of Coval's voice benchmarks, reportedly
On Coval's voice-AI leaderboard Nari reportedly takes #1 in TTS accuracy (WER) and #2 in latency while being the cheapest endpoint, with ASR at lowest latency and #2 accuracy, 0.1% off first. These are vendor-run numbers, so wait for independent runs before betting a product on them.
The inference engine is open source
Nari open-sourced the stack behind it (nari-labs/nari-qwen3-tts), which holds sub-50ms time-to-first-audio at 10 requests/sec on a single H100. Don't want to pay per character? Run the same speed on your own GPU.
Base weights are Apache-2.0 and clone a voice from 3s
Underneath sit Alibaba's Qwen3-TTS models (0.6B and 1.7B, Apache-2.0): voice cloning from a 3-second clip, voice design from a text description, streaming at ~97ms end-to-end, across 10 languages. Qwen3-ASR is open too.
Build this weekend: a voice agent that costs cents
At $5 per 1M characters and $0.06 per hour, a mic-in / ASR / LLM / TTS-out assistant finally pencils out for a side project — or ship it fully offline on the 0.6B weights with no per-call bill at all.