Qwen-Audio-3.0-TTS (Qwen TTS): the #1 speech model you can't download

You send text and a voice choice, it returns speech. What sets this generation apart is control granularity. Tongyi ships 86 inline tags you can drop into the text to trigger emotion, delivery style and non-verbal sounds like laughter or an audible breath, and on top of that you can steer the voice with a plain-language style instruction rather than fiddling with numeric knobs. It also does voice cloning from a reference sample, and Tongyi specifically claims it holds up when the reference audio is not clean, which is the usual failure point for cloning.

The two tiers split the market the way you would expect. Flash is tuned for real-time interaction, quoting first-packet latency around 300ms, which is the range you need for a live voice assistant or a companion that talks back without an awkward pause. Plus trades some of that speed for the top-of-arena quality. Under the hood Tongyi describes a five-stage progressive training pipeline and a 12.5 Hz low-frame-rate speech tokenizer, and it can synthesise up to roughly three minutes of audio in a single pass rather than stitching short clips.

Language coverage is broad: Arabic, Chinese, English, French, German, Indonesian, Italian, Japanese, Korean, Malay, Portuguese, Russian, Spanish, Tagalog, Thai and Vietnamese, plus 20 Chinese dialect regions. Tongyi reports best-in-class error rates (WER/CER) in 10 of those 16 languages, so the multilingual claim is not just a language list, it is backed on the accuracy side for most of them.

At launch, Qwen-Audio-3.0-TTS-Plus took the number one spot on the Artificial Analysis Speech Arena for provider voices at roughly 1,236 Elo, narrowly ahead of Simba 3.2 at about 1,234 and above Gemini 3.1 Flash TTS and Sonic 3.5. That is a real result from an independent blind-preference arena, not a vendor self-report, which is why we lead with it.

The honest footnote: a two-point Elo gap at the very top sits inside overlapping confidence intervals, so calling it a clean win over Simba 3.2 overstates it. It is more accurate to say Qwen is in a statistical tie for first. It is unambiguously in the top group; whether it is strictly the best voice is not something an arena at that margin can settle.

There is also a throughput trade to know about before you build on it. Reporting around launch put Qwen's generation speed well below some rivals on characters per second, so for high-volume batch narration the Plus tier can be slower to grind through a long script than a faster competitor, even though per-clip quality is higher. For interactive use the Flash tier and its 300ms first packet are the relevant numbers; for bulk offline work, measure throughput on your own text before committing.

Pricing is quoted per character, which is hard to feel, so here is the translation. At about $27.59 per one million characters, and taking normal speech at roughly 150 words per minute and about five characters per word, a minute of narration is on the order of 750 characters, which works out to roughly two US cents per minute. A ten-minute explainer voiceover is around twenty cents; an hour of audiobook-style narration is somewhere near a dollar twenty in raw synthesis cost.

In hosted-TTS terms that is cheap: reporting around launch put it at roughly a third of what the incumbents charge for the tiers it outranks. The number to keep in mind is that this is a per-use meter that never stops. Every regeneration, every tweak, every A/B of two reads bills again. For a product doing thousands of minutes a day, the arithmetic changes, and that is where the hosted-versus-local question below stops being philosophical.

One caveat we apply to every vendor figure: the ~$27.59 per million characters is Alibaba's published price, and cloud pricing carries the usual asterisks (region, tier, free-quota changes). Treat it as the right order of magnitude, and check Model Studio's current rate card before you budget a real workload against it.

Good fits: a multilingual product that needs one voice engine across 16 languages without wiring up a different vendor per market; a real-time voice agent or companion where Flash's ~300ms first packet keeps conversation natural; polished narration where the 86 inline tags and natural-language style control earn their keep over a flatter open model. If quality per clip is the thing you are optimising and you are already on Alibaba Cloud, this is a strong default.

Weaker fits: anything where you cannot or will not send text to a third-party cloud (regulated data, private companion transcripts, offline deployments); high-volume batch work where the per-character meter and the slower throughput both bite; and any project whose whole premise is running locally with no account. For those, an open-weight model you host yourself is the honest recommendation, and we would rather point you there than pretend the hosted lock-in does not matter.

If you are choosing an audio stack more broadly rather than just a TTS voice, note that TTS is one piece. Music and full-song generation is a different tool (our Suno model and cost page covers that side), and a companion product usually needs a voice engine, a music engine and sometimes a speech-to-text layer, not one model for all three.