Nari Labs Open-Sources Qwen3 Speech Inference Engine, Claims Top Accuracy on Voice AI Benchmark

The post was written by a person identifying himself as Toby from Nari Labs, who introduced the release as "Show HN: Nari Qwen3-TTS and Qwen3-ASR." The inference engine is published on GitHub at github.com/nari-labs/nari-qwen3-tts. Toby wrote that the company believes the continued dominance of closed-source speech models is an inference problem rather than a model-quality problem, arguing that existing serving systems such as vLLM and SGLang are not well suited to multimodal inference.

According to the company's account of the Coval voice AI benchmark, Nari Labs' Qwen3-TTS endpoint ranks second in latency and first in accuracy as measured by word error rate, ahead of 11Labs, Cartesia and other unnamed competitors, while being the cheapest endpoint on the list. Coval is described in the post as a Y Combinator S24 company. The company also said its Qwen3-ASR endpoint has the lowest latency in the comparison and the second-highest accuracy, trailing the leader by 0.1 percent, and is the second cheapest model listed. These figures come from Nari Labs' own description of the benchmark; the full results were not reproduced in the post.

Nari Labs said its endpoints compare favorably with Alibaba's own official endpoints for the same Qwen3 speech models. "Interestingly, Alibaba's official endpoints seem to perform worse in terms of accuracy and latency compared to ours," Toby wrote, adding that the company still credited the Qwen team for releasing the speech models as open source. The post did not give measurements for the Alibaba endpoints, and Alibaba has not responded publicly to the comparison.

Nari Labs previously released Dia, which it describes as the first open-source text-to-speech model capable of producing natural dialogue. Toby wrote that many strong speech models have been released publicly since then, but that the market remains dominated by closed-source offerings.

According to the post, reaching the published latency and cost figures required "a lot of clever inference engineering" rather than changes to the underlying models. Nari Labs said it intends to keep pushing prices down so that speech synthesis and recognition become commodity capabilities that apps can adopt without worrying about per-unit cost. The company said it is also working on other parts of audio, including speaker diarization, as well as inference for video and world models.

At the time of the post, the item showed 7 points and 1 comment on Hacker News.