It can speak 20 regional Chinese dialects

This is the easiest one for Chinese readers to judge for themselves, since you can usually tell right away when something sounds off. Give it a dialect reference clip, and it uses that same voice to read a different dialect passage.

The 20 dialect regions on the launch page span this range:

It can still clone a voice from a noisy reference clip

In the real world, reference audio rarely comes at studio quality. Record something on your phone in an office and you pick up AC hum and keyboard clatter; record it in a conference room and the room's echo sits on top of the voice. Both used to wreck the cloning results.

The old fix was to add a pre-processing step: run noise reduction to clean up the reference clip before feeding it into the cloning pipeline. The problem is that noise reduction erases timbre detail along with the noise — the cleaner it scrubs, the more the voice sounds sanded down, and the clone ends up sounding less like the original speaker, not more.

The difference is audible in this next set. The reference clip is a noticeably noisy Chinese recording, with three outputs from the model itself and a comparison model.

The launch page tests three kinds of degraded input: noise, reverb, and muffled reference audio (like phone-call bandwidth with the high frequencies cut off).

You can direct it with plain-language instructions

Controlling tone used to mean tweaking a pile of acoustic parameters — pitch, speed multiplier, energy envelope. This generation gives you two layers of control, neither of which touches a parameter.

Describe the effect you want in a sentence

Instructions can be casual or structured. Samples on the launch page range from everyday phrasing like Please read this gently and slowly, like a bedtime story. to structured setups like Character: female morning-radio music host. Setting: opening the show, upbeat but restrained energy. Style: bright, warm, light rhythm. Pace: slightly fast. It also accepts JSON, for example {"speed": "fast", "volume": "loud", "rhythm": "urgent"}.

Insert a tag to control one specific spot

The instruction sets the overall tone; tags control specific spots. There are two kinds: control tags set an emotion or style that holds from where the tag appears until the next control tag shows up, or until the sentence gets auto-split for running too long; sound-effect tags drop in a single sound effect at that exact position without touching the emotion around it.

[excited]What a beautiful day today![laughing]Let's go out and have fun together!

[serious]Please pay attention to the safety precautions.[excited]Alright, let's get started now!

One limitation worth flagging: tags only work in unidirectional streaming mode — meaning the full text is sent up front and the audio streams back. If you want two-way real-time conversation where audio plays while you're still typing, tags won't work there.

How it gets both fast and stable

The part that makes it fast is carried over from the previous generation; what's actually new this time is the training pipeline. Two separate things to look at.

Slicing one second into 12.5 chunks

Generating speech takes two steps: first chop the audio into small discrete units, then generate those units one chunk at a time, and finally reconstruct the sound wave from them. How densely it's chopped is the frame rate. A higher frame rate means more chunks to generate per second, which means slower generation; drop the frame rate and there are fewer steps, so it's naturally faster. This generation runs at 12.5 Hz — 12.5 chunks per second.

Here's how each frame carries more information: 16 layers of encoding are stacked onto every chunk, trading depth for density, using layer count to catch the detail that a lower frame rate would otherwise drop. This tokenizer isn't new to this release either — it was already in use back in January's Qwen3-TTS, and it was open-sourced alongside that model; its technical report cites a first-packet time of 97 milliseconds.

Making up the loss in five steps

The speed gained by dropping the frame rate has to be paid back through the training pipeline, in content accuracy and voice stability. That pipeline is what's actually new this round. Two components to separate first: the LM decides, from the text, which speech units to say; the FM turns those units back into audible sound waves. One reads the sheet music and picks which notes to play; the other actually plays them.

The five-step arc is "train each piece on its own first, then join them and train together, then circle back and patch each one's specific weak spot." The first two steps lay the foundation; the last three are targeted fixes — the LM's weak spot is mispronunciation and stiff delivery, the FM's weak spot is collapsing on dirty reference audio and not preserving timbre well enough.

Beyond these five steps, the launch page mentions two more supporting capabilities: a reproducible speaker fine-tuning pipeline for adapting the model to a specific voice, and vocoder super-resolution for outputting 48 kHz audio. The launch notes mark 48 kHz output as coming soon; the preset voice library covering all 16 languages is already live, so you can pick a built-in voice without cloning one first.

Where it wins, and where it loses

Benchmarks ran on CV3-Eval across all 16 languages, against MiniMax-Speech-2.8-HD-API, ElevenLabs-v3-API, DotsTTS-SOAR (10-step sampling), VoxCPM2, and the company's own Qwen3-TTS-1.7B-Base open-sourced back in January. Two radar charts — farther from the center is better.

Content accuracy (word error rate; radial scale runs from 16 at the center out to 1.45 at the edge — lower is better, so farther out is better). The thick red line is Qwen-Audio-3.0-TTS, hugging the outer edge on most languages, though the dark blue MiniMax line pushes past it on several axes — the gap between the two is razor-thin and hard to eyeball apart on more than one axis. The purple DotsTTS and yellow-green Qwen3-TTS lines collapse toward the center on several spokes, corresponding to languages they don't support. Image source: launch page.

Speaker similarity (radial scale runs from 60 at the center out to 90.76 at the edge — higher means it sounds more like the original speaker). The red line sits on the outermost edge across all 16 languages. The green ElevenLabs line hugs the innermost black baseline ring the whole way around, suggesting its score sits at or below the chart's display floor; the launch page doesn't specify the test conditions for this metric. Image source: launch page.

Timbre fidelity is where it pulls ahead by the widest margin — a clean sweep across all 16 languages. Pronunciation accuracy isn't a clean win: the launch page's claim is that Flash and Plus combined take the lowest word error rate in 10 of the 16 languages, with the other 6 going to rivals. Worth noting: the chart above only plots a single Qwen-Audio-3.0-TTS line without splitting Flash from Plus, so that 10-versus-6 breakdown can't actually be read off the chart itself.

How it ranks on third-party leaderboards

The two charts above are the releasing team's own benchmarks. The third-party evaluator Artificial Analysis also runs a leaderboard based on blind human pairwise comparisons, scored as Elo ratings. On its July 14 "provider's own voices" sub-leaderboard, Qwen-Audio-3.0-TTS-Plus takes first place.

Top 15 on Artificial Analysis's text-to-speech arena leaderboard (provider's own voices sub-leaderboard). First place at 1236 is only 2 points ahead of second-place Simba 3.2's 1234, and their error bars overlap — the leaderboard itself marks both in the 1-2 rank range. Image source: Artificial Analysis official account, 2026-07-14.

That first place needs some context: 1236 versus second-place Simba 3.2's 1234 is only a 2-point gap, and the two error bars overlap — the leaderboard itself lists both in the 1-2 rank range. Further down are Google's Gemini 3.1 Flash TTS (1214) and Cartesia's Sonic 3.5 (1207). ElevenLabs's Eleven v3 sits at 12th place with 1174.

The same leaderboard also supplies two numbers the launch page never mentions: generation speed and price.

On generation speed it comes in dead last on this list: 16 characters per second, nearly half of Simba's, and about a seventh-and-a-half of Sonic 3.5's. On price it costs more than double Simba, about 50% more than Gemini, but roughly 30% less than Sonic 3.5. So that first-place ranking is bought with quality — it doesn't hold an edge on speed or price.

How you can actually use it right now

Access runs through Alibaba Cloud Bailian, with model names qwen-audio-3.0-tts-flash and qwen-audio-3.0-tts-plus, over a WebSocket protocol. Official sample code is provided in six languages: Python, Java, Go, C#, PHP, and Node.js.

One thing is easy to mix up. Tongyi open-sourced a full Qwen3-TTS lineup back in January, with weights on HuggingFace you can download and run locally; Qwen-Audio-3.0-TTS was not open-sourced alongside this release — searching HuggingFace for that name turns up nothing. Here's how the two generations relate:

If you want to run it locally, what's downloadable right now is still January's lineup. Its 1.7B-CustomVoice version has already crossed 2 million downloads on HuggingFace, and the community has produced various quantized versions in GGUF, MLX, and other formats.

One step the demos earlier in this piece skip over: you can't just drop a reference clip straight into a synthesis request. To use your own voice, you first need to register the audio as a custom voice through the voice cloning feature to get a voice ID, then set voice to that ID at synthesis time. Calling it directly, voice is set to the name of a preset system voice.

Qwen-Audio-3.0-TTS, released July 20, ships in two tiers: Flash chases speed, hitting roughly 300 milliseconds from request to first sound; Plus chases quality. One reference clip can speak across 16 languages (7 of them new, clustered around Southeast Asia and the Middle East), and it also covers 20 Chinese dialect regions, from Cantonese and Shanghainese to Gansu and Qingdao dialects. Noisy reference audio doesn't need pre-cleaning — the model trained on dirty recordings from the start, doing noise suppression and timbre preservation together.

Controlling delivery changed too. Instead of tweaking acoustic parameters, you now write tags straight into the text, and they take effect from wherever they appear:

Alibaba Cloud's docs publish 31 such tags (24 for delivery, 7 for sound effects), while the launch page claims 86 new ones — the gap isn't explained. These benefits only hold up if the model is fast and stable enough, which comes down to the system below.

XiaoHu recorded a short reference clip of himself speaking Chinese and wanted it to speak Japanese instead. The model's approach: chop the audio into small chunks, generate them one chunk at a time, then reconstruct the sound wave. How many chunks per second determines how many steps it has to generate — that's the master switch for speed.

Turn "characters per second" into something concrete: how long each service takes to read out a 3,000-character piece.