Nari Qwen3-TTS and Qwen3-ASR Performance Benchmarks

Nari Labs has achieved leading positions on the Coval voice AI benchmarks as of September 14, 2026, placing its models on the quality-latency Pareto frontier for both Text-to-Speech (TTS) and Speech-to-Text (STT). The company reports leading the latency-cost and quality-cost Pareto frontiers among all publicly available models on the benchmark.

Speech-to-Text (ASR) Performance

The Nari Qwen3-ASR Fast model ranks first in Time-to-Final-Segment (TTFS) with a p50 of 44 ms and a Word Error Rate (WER) of 3.6%. In terms of accuracy, it ranks second, trailing AssemblyAI’s Universal 3.5 Pro (3.5% WER).

Cost Efficiency

Qwen3-ASR Fast is priced at $0.12 per hour, tying for the second-lowest price among models with known public rates in the Coval pricing directory. This represents a significant cost reduction compared to competitors:

• Universal 3.5 Pro: 3.75× more expensive
• Deepgram Nova 3: 2.4× more expensive
• Nari Standard endpoint: $0.06 per hour

Text-to-Speech (TTS) Performance

The Nari Qwen3-TTS Fast model ranks first in Word Error Rate (WER) at 3.8% and second in Time-to-First-Audio (TTFA) with a p50 of 63 ms. The only model with lower latency is Fluxions' vui (49 ms), which is a 300M parameter model compared to the 1.7B parameter Qwen3-TTS model served by Nari.

Cost Efficiency

At $10 per 1 million characters, the Qwen3-TTS Fast endpoint is tied for the cheapest model in the Coval pricing directory. This compares to:

• ElevenLabs Eleven v3 Conversational: 5× more expensive
• Cartesia Sonic 3.6: 6.5× more expensive
• Nari Standard endpoint: $5 per 1 million characters

Infrastructure Comparison

Nari Labs reports significantly better performance using the same model as the official Qwen3 TTS Flash Realtime endpoint, which records a 692 ms median TTFA and 8.8% WER. Nari also outperforms Baseten’s dedicated Qwen3-TTS endpoint, which has a 101 ms median TTFA and 6.0% WER.

Community Feedback and Technical Discussion

Community reactions on Hacker News highlight both the technical promise and the practical challenges of high-performance voice AI.

Technical Observations

Users noted issues with voice consistency and playback speed, with one user stating:

All TTS generations are too fast. It's almost I'm listening to a podcast on 1.25-1.5x speed.

Another user reported a voice switch occurring midway through a 33-second clip.

Market and Evaluation Perspectives

Industry participants emphasized the need for independent verification and the importance of diverse model options:

• Independent Evals: Users suggested that third-party verification from entities like Datapoint AI is necessary to validate quality claims.
• Market Dynamics: A developer relations representative from AssemblyAI noted that voice models are not a "winner take all market" unlike LLM APIs.
• Future Directions: Some users suggested that the next major evolution will be the creation of cheaper, open-source alternatives to integrated real-time voice experiences like GPT-Live-1.