Alibaba’s Tongyi Lab Releases Qwen-Audio-3.0-TTS, a Hosted Text-to-Speech Model in Flash and Plus Tiers Across 16 Languages
Alibaba’s Tongyi Lab has released Qwen-Audio-3.0-TTS, a production-oriented text-to-speech (TTS) system. The model ships in two variants from the same lineage. Flash targets real-time interaction. Plus targets high-quality generation. Both are delivered as hosted models through Alibaba Cloud Model Studio, not as downloadable weights.
The release focuses on four things developers hit in production: broader language coverage, natural-language style control, fine-grained tag control, and robustness when the reference audio is not clean. Qwen-Audio-3.0-TTS-Plus also ranks first on the independent Artificial Analysis Text-to-Speech leaderboard.
Two variants, one lineage
The two tiers map to different jobs. Flash is tuned for real-time interaction, with first-packet latency at the 300 ms level. Plus is tuned for high-quality generation, where naturalness and timbre fidelity matter more than speed.
The model IDs are qwen-audio-3.0-tts-flash and qwen-audio-3.0-tts-plus. Both are called over a bidirectional WebSocket streaming protocol. The API supports PCM, WAV, MP3, and Opus, with sample-rate output up to 48 kHz. It exposes streaming input and output, voice cloning, Voice Design, and instruction control. Alibaba provides the DashScope SDK plus raw WebSocket examples in Python, Java, Go, C#, PHP, and Node.js, across its Singapore and Beijing regions.
How the model is built
Two design choices anchor the system.
- A 12.5 Hz low-frame-rate speech tokenizer reduces autoregressive decoding cost while retaining content and speaker information. A lower frame rate means fewer tokens per second of audio, which cuts inference latency.
- A five-stage progressive training paradigm coordinates the language model (LM) and flow-matching (FM) components. The stages are independent LM and FM pretraining, joint training with high-quality data annealing, LM reinforcement learning, FM robustness training, and FM reinforcement learning. The research team reports this pipeline improves content consistency, prosodic naturalness, voice fidelity, perceptual quality, and robustness.
The model also handles one-pass long-form synthesis up to 3 minutes, hard text-normalization cases, and vocoder super-resolution for 48 kHz output.
