Show HN: Nari Qwen3-TTS and Qwen3-ASR – High accuracy, low latency and cost
summary
Nari Labs has achieved top rankings in Coval's voice AI benchmarks for both Speech-to-Text and Text-to-Speech technologies. Their models maintain a superior balance of low latency and high accuracy, positioning them on the quality-latency Pareto Frontier. This performance is coupled with aggressive pricing, making their tools more cost-effective than competing high-performance models. The company aims to provide highly responsive, low-cost AI voice agents for production use.
key points
- The Qwen3-ASR Fast model leads the industry in latency for Speech-to-Text, recording a p50 of 44 ms while maintaining a competitive Word Error Rate. It is significantly cheaper than competitors like AssemblyAI and Deepgram.
- The Qwen3-TTS Fast model ranks second in latency and first in accuracy for Text-to-Speech, outperforming official Qwen3 endpoints. It is currently tied for the cheapest available model in the Coval pricing directory.
- Coval's benchmarks focus on critical metrics like Time-to-First-Audio and Time-to-Final-Segment, which are essential for making voice AI feel natural and responsiveness.
who should read it
Developers and engineers building voice AI agents who need to optimize for speed, cost, and accuracy.
discussion positive
consensus
Commenters generally view the new audio models as a significant step forward in the efficiency and quality of open-source TTS and ASR. While there is enthusiasm for the cost-effectiveness and speed of the approach, some users report technical issues like unnatural speaking speed and voice switching. There is a general agreement that the TTS market is fragmented and will remain competitive rather than being dominated by a single provider.
top insight
The most valuable perspective is that TTS models are not a 'winner-take-all' market like LLM APIs, meaning multiple specialized providers can coexist and thrive. This suggests that the scale of the efficiency gains provided by the Nari Labs approach is more important than achieving absolute market dominance. The discussion emphasizes that the technical approach to inference speed is a key lever for accessibility.
Generated by an LLM from the article and thread. Check the sources before relying on it.