Nari Qwen3 Speech: High-accuracy, low-latency text-to-speech and ASR models

What it does

Nari Qwen3 Speech provides open-source text-to-speech and automatic speech recognition models with an inference engine optimized for speed and accuracy. The company built specialized infrastructure to run these models faster than general-purpose inference systems like vLLM and SGLang.

Who it is for

Developers building voice applications, AI agents, and voice interfaces need low-latency speech processing. Teams evaluating alternatives to closed-source speech APIs from providers like ElevenLabs, Deepgram, and AssemblyAI would find this relevant.

Pricing

Nari offers tiered endpoints. The Fast speech-to-text endpoint costs $0.12 per hour, with a Standard endpoint at $0.06 per hour. For text-to-speech, the Fast endpoint is $10 per 1 million characters, and Standard is $5 per 1 million characters.

How it stands out

According to Coval's benchmark (as of September 2026), Nari ranks first in speech-to-text latency at 44 milliseconds and first in text-to-speech word error rate at 3.8%. The company claims tied-lowest pricing for speech-to-text and tied-lowest for text-to-speech compared to known public rates. Nari serves the same Qwen3 model as official endpoints but reports substantially lower latency and error rates, suggesting their inference optimization delivers meaningful advantages. The text-to-speech model outputs first audio in 63 milliseconds median latency.

What a founder should check

Firstly, verify whether the latency and accuracy advantages persist beyond the benchmark window. Coval's benchmarks update every 30 minutes, and competitors may improve their implementations. Check if this performance advantage is durable or temporary.

Secondly, examine switching costs for existing customers. ElevenLabs, Deepgram, and AssemblyAI have entrenched integrations and customer relationships. Measure what friction a developer faces migrating from incumbents—API differences, data portability, retraining workflows.

Thirdly, assess the moat around Nari's inference engine. The company open-sourced the codebase, which means competitors can study and replicate the approach. Determine whether Nari's advantage depends on proprietary hardware access, specialized expertise, or continuous engineering that is harder to copy.

Thinking of building something like this?

Every launch here is a competitor to somebody's idea. If yours is close, check it against the market before you build: the Full Check names the rivals, the prices and the gaps.