Nari Qwen3-TTS and Qwen3-ASR review (2026): honest scoring on latency, accuracy, pricing, and the publishing gap for creators who need finished content.
Nari Qwen3-TTS and Qwen3-ASR are, on the September 2026 Coval benchmark, about the best combination of latency, accuracy, and price you can rent for voice: sub-70ms first-sound and first-response latency, low word-error rates, and pricing well under premium APIs. As voice and speech-recognition models they earn a high score. The caveats are scope and framing — the headline latency is built for live voice agents, not async content, and the models make audio and transcripts, not finished, published posts. Score them as the excellent audio building blocks they are, not as a content tool.
Nari Labs earned its reputation with Dia, an open-source dialogue TTS model released under Apache 2.0 and covered as a challenger to incumbents like ElevenLabs. In September 2026 the team turned that engineering toward serving Alibaba's Qwen3 speech models fast and cheap, and the result landed at the top of Coval's public voice-AI benchmark — first in speech-to-text latency and text-to-speech accuracy, second in the reverse of each. That is a genuinely strong showing, and this review scores it as one.
I score these for what they are: hosted voice and speech-recognition models. Qwen3-TTS 1.7B turns text into speech; Qwen3-ASR 1.7B turns speech into text. They are not a content-creation suite, and I don't grade them as one — no captions, no video, no images, no scheduling, no publishing. Where they compete, against other TTS and ASR providers, they compete at the front of the pack on the axes that benchmark measures.
Two things anchor the verdict. First, the numbers are real and the price is aggressive: on the benchmark, Nari's TTS reported roughly 63 ms time-to-first-audio at about 3.8% WER, its ASR roughly 44 ms time-to-final-segment at about 3.6% WER, and pricing sits at $10 per 1M characters (TTS Fast) and $0.12 per hour (ASR Fast) — well below the incumbent voice APIs. Second, the honest limits: the latency that tops the chart is a live-voice-agent metric, most of the model's real-world details (voices, languages, cloning) inherit from the Qwen3 base and are worth confirming, and everything downstream of the audio file — the video, the copy, the distribution — is yours to build or bring.
Everything below reflects the state at the September 14, 2026 benchmark and is verified against Nari's announcement and Coval's benchmark. Voice benchmarks re-run on a schedule and provider pricing changes, so confirm current numbers before you commit.
Nari Qwen3-TTS and Qwen3-ASR are two hosted models from Nari Labs, each an optimized serving of one of Alibaba's Qwen3 speech models on Nari's own low-latency inference stack. Qwen3-TTS 1.7B is the text-to-speech model — it converts written text into natural, streaming speech. Qwen3-ASR 1.7B is the automatic-speech-recognition model — it transcribes spoken audio into text. Both are offered on a "Fast" and a "Standard" tier, trading a little latency for a lower price on Standard. Nari's engineering is the differentiator. In August 2026 the team open-sourced, under Apache 2.0, an inference engine for Qwen3-TTS 1.7B that reached sub-50 ms p95 time-to-first-audio at 10 requests per second on a single NVIDIA H100 — using techniques like a unified scheduler, dynamic leading-silence trimming, state-cached incremental decoding, and CUDA graphs. That same discipline is what puts the hosted models on the quality-latency Pareto frontier on Coval's benchmark. For reference, the benchmark measured the official Qwen3 TTS Flash realtime endpoint at 8.8% WER and 692 ms median time-to-first-audio, so Nari's serving is both more accurate and far faster to first sound. The models are what they are — voice and transcription. They ship an audio file or a transcript over an API, and everything around that is left to the developer or creator.
These models fit developers and product teams building real-time voice first: phone agents, live assistants, in-app narration, transcription pipelines — anywhere sub-100 ms first-sound latency and low per-unit cost decide whether the experience feels instant and whether the economics work at volume. They also fit cost-sensitive creators and teams who need a lot of narration or a lot of transcription and want to stop paying premium-API rates, and self-hosters who want Nari's Apache-2.0 TTS engine on their own H100. Where they fit poorly is anyone expecting a finished content tool: these models make audio and transcripts, not captioned video, carousels, blogs, or published posts, and they have no brand-voice-across-a-week layer, no scheduler, and no app. If your bottleneck is producing and distributing finished content rather than synthesizing speech or transcribing it, this is one component, not the pipeline.
| Dimension | Score | Why |
|---|---|---|
| TTS accuracy (word error rate) | 4.6 / 5 | First in accuracy on the September 2026 Coval benchmark at ~3.8% WER — well ahead of the official Qwen3 Flash realtime endpoint's 8.8%. |
| ASR accuracy (word error rate) | 4.4 / 5 | Second in accuracy at ~3.6% WER — a low enough error rate for transcripts you can repurpose with light cleanup. |
| Latency / real-time responsiveness | 4.8 / 5 | The standout: ~63 ms time-to-first-audio (TTS) and ~44 ms time-to-final-segment (ASR), topping the latency charts — built for live, interruptible voice. |
| Pricing & value | 4.6 / 5 | TTS at $10 / 1M chars (Fast) and $5 (Standard); ASR at $0.12 / hr (Fast) and $0.06 (Standard) — Nari cites roughly 3.75–6.5x cheaper than premium voice APIs. |
| Openness / self-host option | 4.0 / 5 | Nari's Qwen3-TTS inference engine is Apache-2.0 and self-hostable (~$2 / 1M chars at full utilization on one H100); the hosted API covers the rest. |
| Developer integration | 4.0 / 5 | Streaming and non-streaming over HTTP plus WebSocket input streaming; a clean fit for real-time apps, though it is an API, not a no-code product. |
| Voice range, languages & cloning | 3.5 / 5 | These inherit from the Qwen3 speech base rather than being Nari-specific; strong on paper but worth confirming voices, language coverage, and any cloning on Nari's docs. |
| Content-workflow scope | 1.5 / 5 | Audio and transcripts only — no written content, images, captions, carousels, scheduling, or publishing. Not what these models are for. |
On price, these models are hard to argue with. Nari lists text-to-speech at $10 per 1M characters on the Fast tier and $5 per 1M on Standard, and speech-to-text at $0.12 per hour of audio on Fast and $0.06 on Standard. Against premium voice APIs — Nari cites ElevenLabs and Cartesia TTS pricing at roughly 5–6.5x higher per character — that is a structural cost advantage for anyone generating a lot of narration or transcribing a lot of audio. At $0.06–$0.12 an hour, transcribing everything you record stops being a budget decision, which is quietly the most useful thing here for a content operation.
The self-host path sharpens the value further. Because Nari open-sourced its Qwen3-TTS inference engine under Apache 2.0, a team with its own H100 can run TTS at roughly $2 per 1M characters at full utilization — trading the convenience of a hosted API for near-marginal-cost audio at scale. That is a real option for high-volume, cost-sensitive workloads, though it comes with the usual engineering and maintenance burden of running your own inference.
The read: as voice and speech models, Nari's Qwen3 servings are priced and benchmarked at the front of the market. What that price does not include is any of the content-production work around the audio — writing the copy, making the visuals, assembling the video, or publishing anything. That is not a criticism; it is scope. You are getting excellent, cheap audio and transcription, and only that.
| Use case | Fit | Why |
|---|---|---|
| Real-time voice agents (phone bots, live assistants) | Strong | Sub-70 ms first-sound and first-token latency at low cost is exactly what a live, interruptible voice experience needs. |
| High-volume transcription of podcasts, calls, or video | Strong | Low WER at $0.06–$0.12 per hour makes transcribing everything you record economical. |
| Cheap narration for faceless or long-form audio | Strong | Low per-character TTS pricing suits batch voiceover for articles, explainers, and audio versions. |
| Self-hosting a fast TTS engine at scale | OK | The Apache-2.0 Qwen3-TTS engine gets you to ~$2 / 1M chars, but only if you can run and maintain an H100 inference stack. |
| Generating written posts, scripts, or blogs | Weak | These models read and transcribe; they do not write copy, and there is no brand-voice layer. |
| Producing visual content (carousels, quote cards, video) | Weak | They output audio and text transcripts; no images or video assets. |
| Scheduling and publishing across platforms | Weak | No scheduler and no social connections — these models publish nothing. |
To be clear where I stand: I run Kompozy, and Kompozy is not a competitor to these models. Nari's Qwen3-TTS and Qwen3-ASR are voice and speech models you call over an API. Kompozy is a content operation you run. I include this note because a fair number of people find benchmark-topping voice models while trying to solve a content-volume problem, drawn in by the low price — and it is worth saying plainly that a faster, cheaper voice won't solve that.
The tell is the metric everyone quotes. Sub-70 ms time-to-first-audio is a live-voice-agent number — it matters when a human is waiting on the line for a bot to answer. Async content creation doesn't care whether a voiceover starts in 63 ms or 700 ms; it renders in the background either way. So for a creator, the real draw here is cost and accuracy, not latency — and even then, what you get is an audio file or a transcript, not a post. If your bottleneck is turning one idea into a week of on-brand posts across nine platforms — copy under a Persona Brief, short-form and avatar video (with its own built-in TTS), carousels, quote cards, a blog, and a newsletter, scheduled and published from one queue — that is a content engine's job, and it is the job Kompozy is built for. The clean pairing many teams land on: cheap Qwen3-ASR to transcribe what you record, Kompozy to turn that transcript into the finished, published content — two tools, two halves, and the audio half costs almost nothing.
For voice and transcription, yes. On Coval's September 2026 benchmark they lead on latency and accuracy at pricing well under premium APIs, which is a strong package for real-time voice or high-volume transcription. They are not worth it as a content tool, because they generate no written posts, images, or video and publish nothing — they are audio and speech models, not a workflow.
On the September 14, 2026 Coval benchmark, Nari's TTS reported about 3.8% word error rate (first in accuracy) and its ASR about 3.6% (second). For reference, the benchmark measured the official Qwen3 TTS Flash realtime endpoint at 8.8% WER. Treat exact figures as a benchmark-day snapshot and confirm on Coval's page.
Nari lists TTS at $10 per 1M characters (Fast) and $5 per 1M (Standard), and ASR at $0.12 per hour (Fast) and $0.06 per hour (Standard). Self-hosting Nari's Apache-2.0 Qwen3-TTS engine works out to roughly $2 per 1M characters at full utilization on one H100. Confirm current pricing on Nari's site.
On the benchmark axes — latency, TTS accuracy, and price — Nari leads, and it is far cheaper per character. ElevenLabs is a broader audio suite with strong voice cloning, wide language coverage, and licensed music. If you want the cheapest fast voice or transcription, Nari is compelling; if you want a full audio toolkit, ElevenLabs covers more. Benchmark both on your own scripts.
Yes for TTS. Nari open-sourced its Qwen3-TTS 1.7B inference engine under Apache 2.0; it reached sub-50 ms p95 time-to-first-audio at 10 requests per second on a single H100 and works out to roughly $2 per 1M characters at full utilization. That trades hosted convenience for near-marginal-cost audio, with the burden of running your own inference.
No. They generate audio (TTS) or transcripts (ASR) only — no posts, images, video, captions, or scheduling. To turn a cheap voiceover or a transcript into finished, on-brand posts across platforms you need a content engine like Kompozy, which many teams pair with a low-cost ASR or TTS model.
Developers building real-time voice (agents, assistants, live narration), teams transcribing a lot of audio cheaply, and self-hosters who want a fast, open TTS engine. They fit poorly for anyone whose real need is producing and distributing finished, on-brand content across platforms.
See Nari Qwen3-TTS & Qwen3-ASR vs Kompozy comparison → · Get Started →