Alibaba’s Tongyi Lab Releases Qwen-Audio-3.0-TTS, a Hosted Text-to-Speech Model in Flash and Plus Tiers Across 16 Languages
MarkTechPost Asif Razzaq
Alibaba's Tongyi Lab dropped Qwen-Audio-3.0-TTS, a cloud-only text-to-speech model in fast and high-quality flavors. It just topped an independent voice leaderboard while charging way less than rivals.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Alibaba's Tongyi Lab has quietly put out a text-to-speech system that's already causing a stir, and not because of some flashy demo. Qwen-Audio-3.0-TTS comes in two tiers, Flash and Plus, both accessed only through Alibaba Cloud Model Studio as hosted APIs. No weights to download here. Flash is built for real-time use, hitting first-packet latency around 300 milliseconds, while Plus trades speed for polish, aiming at naturalness and voice fidelity over raw responsiveness.
Under the hood, the team leaned on a 12.5 Hz low-frame-rate speech tokenizer, which basically means fewer tokens per second of audio and cheaper, faster decoding. Getting there took a five-stage training process that separately pretrains the language model and flow-matching components before joining them together, adding reinforcement learning passes along the way. Tongyi Lab says this pipeline is what pushed up content consistency, prosody, and robustness, and the model can now handle up to three minutes of continuous synthesis in one pass along with tricky text-normalization cases.
Language coverage is where this release earns its keep. Sixteen languages total, seven of them new additions, plus 20 Chinese dialect regions. Flash posts the lowest average error rate at 3.87, Plus trails slightly at 3.96, and together the family claims the best word or character error rate in 10 of the 16 languages tested. On speaker similarity, Plus leads across all 16 languages with an average score of 82.75. There's also a preset voice library now, so developers don't have to clone a voice from scratch just to ship something usable.
Control is handled through 86 inline tags dropped straight into the text, split between ones that shift tone until the next tag, like [excited] or [whispers], and ones that insert a single effect, like [laughing] or [clears throat]. It's a clever, granular way to hint at emotion without building a separate control layer, though the catch is these tags only work in unidirectional streaming mode, not the bidirectional kind.
The headline number, though, is the leaderboard spot. Qwen-Audio-3.0-TTS-Plus took the top quality ranking on the Artificial Analysis Speech Arena for Provider Voices, scoring an Elo near 1,236, just ahead of Simba 3.2 at 1,234 and clear of Gemini 3.1 Flash TTS and Sonic 3.5. That gap over Simba 3.2 is small enough to fall inside overlapping confidence intervals, so call it a statistical tie at the top rather than a knockout. And the trade-offs are real: Plus generates roughly 16 characters per second, well behind competitors like Sonic 3.5 at 120, even as it undercuts ElevenLabs and MiniMax pricing by a wide margin at $27.59 per million characters.
My take — AI-written commentary, not fact-checked reporting
A hosted-only model topping a leaderboard while charging a fraction of the going rate is the kind of thing that should worry incumbents more than it seems to right now. Sure, the throughput numbers are rough and the naming clash with the open Qwen3-TTS line is a minor headache, but pricing pressure like this tends to force everyone else's hand eventually. The real story isn't the Elo score, it's that a non-Western player just made the premium TTS tier look overpriced.
Read more about this at: MarkTechPost