Alibaba Qwen-Audio-3.0-TTS: 16 Languages, Two Tiers, $27.59 per 1M Chars
Alibaba's Qwen-Audio-3.0-TTS ships in Flash and Plus tiers across 16 languages at $27.59/1M chars. Plus tops the Artificial Analysis TTS leaderboard at ~1,236 Elo.
Alibaba's Tongyi Lab released Qwen-Audio-3.0-TTS on July 20, 2026, a hosted text-to-speech system in two tiers: Flash, tuned for real-time interaction with a first-packet latency around 300 milliseconds, and Plus, optimised for voice quality and timbre fidelity. Both are available exclusively through Alibaba Cloud Model Studio, with no downloadable weights. Plus debuted at the top of the Artificial Analysis Speech Arena at approximately 1,236 Elo and is priced at $27.59 per one million characters, well below several established rivals.
What happened
| Detail | Value |
|---|---|
| Release date | July 20, 2026 |
| Model IDs | qwen-audio-3.0-tts-flash / qwen-audio-3.0-tts-plus |
| Languages supported | 16 languages, 20 Chinese dialect regions |
| Flash first-packet latency | ~300 ms |
| Plus Elo (Artificial Analysis arena) | ~1,236 |
| Price | $27.59 per 1M characters |
| Plus throughput | ~16 characters per second |
| Max audio output quality | 48 kHz |
| Max single-pass synthesis | 3 minutes |
| Fine-grained inline tags | 86 |
Tongyi Lab built the system around two architectural choices. First, a 12.5 Hz low-frame-rate speech tokenizer cuts autoregressive decoding costs: fewer tokens per second of audio means lower latency. Second, a five-stage progressive training pipeline coordinates the language model and flow-matching (a technique for generating smooth, natural-sounding audio) components, running from independent pretraining through two reinforcement learning stages aimed at robustness and perceptual quality.
The API streams over a bidirectional WebSocket protocol and supports PCM, WAV, MP3, and Opus formats. Alibaba provides SDKs in Python, Java, Go, C#, PHP, and Node.js, served from Singapore and Beijing regions.
How does the quality compare to rivals?
| Model | Elo (arena) | Throughput (chars/sec) | Price per 1M chars |
|---|---|---|---|
| Qwen-Audio-3.0-TTS-Plus | ~1,236 | ~16 | $27.59 |
| Simba 3.2 | ~1,234 | 30.2 | N/A in source |
| Gemini 3.1 Flash TTS | 1,214 | 27 | N/A in source |
| Sonic 3.5 | 1,207 | 120 | N/A in source |
One important caveat: the gap between Plus and Simba 3.2 falls inside overlapping confidence intervals on the leaderboard. Treat that top ranking as a statistical tie rather than a clear win. The quality lead over third and fourth place is more meaningful.
On multilingual accuracy, measured as word/character error rate (lower is better), Flash posts a 3.87 average across 16 languages and Plus sits at 3.96. The model family leads in 10 of the 16 languages. For speaker similarity, Plus scores 82.75 and Flash 80.44 across all 16 languages.
The 16 supported languages are Arabic, Chinese, English, French, German, Indonesian, Italian, Japanese, Korean, Malay, Portuguese, Russian, Spanish, Tagalog, Thai, and Vietnamese. Seven of those are new versus the previous model line.
Fine-grained control: what the 86 tags do
The release ships with 86 inline tags that developers embed directly in the target text to control delivery at the phrase or word level. Two categories apply:
- Control tags such as [excited], [sad], [whispers], and [asmr] set an emotion or style that persists until the next tag appears.
- Rich-language tags such as [laughing], [gasp], and [clears throat] insert a single vocal event without altering surrounding tone.
A sample string from Tongyi’s documentation: [excited]What a beautiful day today![laughing]Let’s go out and have fun together!
One restriction to note: both tag types work only in unidirectional streaming mode, not bidirectional. That rules them out for some real-time conversational setups.
Why it matters
Most top-ranked TTS systems right now come from Western or established API providers. A non-Western model topping an independent quality leaderboard at a third of the price of ElevenLabs or MiniMax gives businesses with multilingual or cost-sensitive voice requirements a genuinely competitive alternative to evaluate.
The language breadth is also notable. Southeast Asian languages like Tagalog, Malay, Indonesian, and Thai are rarely covered well by the major Western TTS providers. If your product serves those markets, the preset voice library alone (no cloning required) cuts time to ship.
For teams already exploring AI integration in customer-facing products, a hosted TTS API with 48 kHz output, voice cloning, and an SDK in six languages is infrastructure that’s ready to wire in today.
Our take
The price story is real and the leaderboard result is credible, but the throughput number is the thing we’d stress-test first. At roughly 16 characters per second, Plus is the slowest system near the top of the Artificial Analysis arena by a wide margin. Sonic 3.5 delivers 120 characters per second. For batch narration or async content generation, that gap may not matter. For anything approaching live voice UI, it likely will.
The hosted-only model is a practical constraint worth naming. You are dependent on Alibaba Cloud availability and pricing decisions. That’s a normal SaaS trade-off, but it’s different from running Qwen3-TTS locally. The naming overlap between this hosted product and the open-weight Qwen3-TTS line has already caused confusion in developer communities, so double-check which model you’re actually referencing in documentation.
If you’re comparing TTS options for a real project, benchmark throughput under your actual load before price comparison becomes relevant. You can follow broader AI model releases like this one in our AI news coverage.
What to do about it
- Confirm your latency and throughput requirements before choosing Flash vs. Plus. Real-time voice UI needs Flash; high-quality narration or dubbing is the Plus use case.
- Test the 86 inline tags in unidirectional mode if expressive control is a requirement for your project.
- Verify current pricing and regional availability on Alibaba Cloud Model Studio directly. Both price and leaderboard rank change frequently.
- Run a side-by-side quality test with your actual content. Arena Elo scores reflect average human preference, not your specific language pair or voice style.
If throughput matters as much as quality, Sonic 3.5 at 120 characters per second is the benchmark to beat before committing to Plus.
Frequently asked questions
What is the price of Qwen-Audio-3.0-TTS-Plus?
Alibaba lists the rate at $27.59 per one million characters, which is roughly one-third of what ElevenLabs and MiniMax charge for comparable tiers, according to the Tongyi Lab release.
How does Qwen-Audio-3.0-TTS rank compared to other TTS models?
Qwen-Audio-3.0-TTS-Plus topped the Artificial Analysis Speech Arena for Provider Voices with an Elo of approximately 1,236, just ahead of Simba 3.2 at 1,234. The gap falls within overlapping confidence intervals, making it a statistical tie at the very top.
What languages does Qwen-Audio-3.0-TTS support?
The model supports 16 languages: Arabic, Chinese, English, French, German, Indonesian, Italian, Japanese, Korean, Malay, Portuguese, Russian, Spanish, Tagalog, Thai, and Vietnamese. It also covers 20 Chinese dialect regions, with seven languages newly added versus the prior model line.
What is the throughput of Qwen-Audio-3.0-TTS-Plus and is it fast enough for real-time use?
Plus generates approximately 16 characters per second, which is significantly slower than rivals like Sonic 3.5 (120 chars/sec) and Simba 3.2 (30.2 chars/sec). For real-time conversational use cases, the Flash tier with its ~300ms first-packet latency is the intended option.
