Show HN: Nari Qwen3-TTS and Qwen3-ASR – High accuracy, low latency and cost
Show HN: Nari Qwen3-TTS and Qwen3-ASR – High accuracy, low latency and cost
TL;DR: Nari Labs has secured a leading position in Coval’s voice AI benchmarks, occupying the quality-latency Pareto Frontier for both Speech-to-Text (STT) and Text-to-Speech (TTS). Furthermore, they dominate the latency-cost and quality-cost frontiers among all public models.
The Benchmark Standard: Coval
Coval is a premier authority in the evaluation of voice AI, providing the industry with some of the most respected benchmarks to help developers optimize speech agents for production environments.
Key Performance Indicators (KPIs)
To understand these results, we must define the primary metrics used:
- Text-to-Speech (TTS): Measured by Word Error Rate (WER) and Time-to-First-Audio (TTFA)—the duration from text input to the first audible sound.
- Speech-to-Text (STT): Measured by WER and Time-to-Final-Segment (TTFS)—the duration from the finalization of a user request to the resulting text.
Low latency (TTFA/TTFS) is essential to prevent voice agents from feeling sluggish, while a low WER is the baseline for accuracy.
🎙️ Speech-to-Text (STT) Performance
As of mid-September 2026, Nari Labs is at the top of the STT charts. Their Qwen3-ASR Fast model is currently the fastest in the industry.
Performance Metrics
| Model | TTFS (p50) | WER | Rank (Latency) | Rank (Quality) |
|---|---|---|---|---|
| Qwen3-ASR Fast | 44 ms | 3.6% | #1 | #2 |
| Universal 3.5 Pro | Higher | 3.5% | Lower | #1 |
Cost Analysis
Nari's pricing is highly competitive. At $0.12 / hour, the Fast endpoint is tied for the second-lowest price in the Coval directory.
- Universal 3.5 Pro: more expensive.
- Deepgram Nova 3: more expensive.
Standard PricingThe Standard endpoint would be the absolute cheapest at$0.06 / hour.
🔊 Text-to-Speech (TTS) Performance
The Qwen3-TTS Fast model demonstrates a superior balance of scale and speed.
The Latency-Quality Tradeoff
Nari Labs ranks #2 in TTFA with a p50 of 63 ms and holds the #1 spot for WER at 3.8%.
While the vui model from Fluxions has a lower median TTFA (49 ms), it is a significantly smaller model:
- Fluxions vui: 300M parameters.
- Nari Qwen3-TTS: 1.7B parameters.
TTS Pricing Comparison
| Provider | Model | Cost (per 1M chars) | Price Multiple |
|---|---|---|---|
| Nari Labs | Qwen3-TTS Fast | $10 | 1x (Tied #1) |
| ElevenLabs | Eleven v3 Conv. | $50 | 5x |
| Cartesia | Sonic 3.6 | $65 | 6.5x |
Note: The Standard endpoint would further reduce costs to $5 per 1M characters.
📊 Comparative Analysis
Nari's implementation significantly outperforms other Qwen3-based deployments.
🚀 Get Started
Nari Labs is transitioning from Public Beta to General Availability (GA) this week.
Onboarding Checklist
- Create an account to claim credits.
- Test the STT and TTS endpoints for free (limited time).
- Check email for the
$20credit grant during the GA switch.
Integration Example
To integrate the API, you might use a configuration similar to this:
{
"model": "qwen3-tts-fast",
"parameters": {
"latency_mode": "ultra-low",
"quality": "high"
},
"endpoint": "api.narilabs.ai/v1/tts"
}
Need help with capacity or latency requirements? Contact the Nari engineering team today.
Footnotes: * Benchmark data captured from Coval’s 1-day view on September 14, 2026, at 15:00 UTC. Note that Coval benchmarks are dynamic and can shift every 30 minutes. * Pricing is based on Nari's published rates compared against known public rates in the Coval directory.