Alibaba’s Tongyi Lab Releases Qwen-Audio-3.0-TTS, Text-to-Speech Model with Flash and Plus Tiers in All 16 Languages
Alibaba’s Tongyi Lab has been released Qwen-Audio-3.0-TTSa production-oriented text-to-speech (TTS) system. The model comes in two versions from the same list. Flash direct real-time interaction. in addition aimed at high quality workmanship. Both are delivered as managed models via Alibaba Cloud Model Studio, not as downloadable weights.
The release focuses on four things that engineers went through in production: comprehensive language coverage, natural language style control, fine-tuned tag control, and robustness when the reference audio is not clean. Qwen-Audio-3.0-TTS-Plus also ranks first on the independent leaderboard for text-to-speech analysis.
Two varieties, one list
The two categories identify different functions. Flash optimized for real-time communication, with a first packet delay of 300 ms. in addition optimized for high-quality generation, where naturalness and fidelity of timbre are more important than speed.
Model IDs are qwen-audio-3.0-tts-flash again qwen-audio-3.0-tts-plus. Both are called the bidirectional WebSocket broadcast protocol. The API supports PCM, WAV, MP3, and Opus, with output sample-rates up to 48 kHz. It covers input and output for streaming, voice synthesis, Voice Design, and command control. Alibaba provides the DashScope SDK and 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 options reinforce the system.
- A 12.5 Hz speech token with low frame rate reduces the cost of automatic recording while preserving the content and information of the speaker. A lower frame rate means fewer tokens per second of audio, which reduces latency.
- A a five-stage training paradigm integrates language modeling (LM) and flow-matching (FM) components. The categories are independent LM and FM pre-training, joint training with high data quality, LM reinforcement learning, FM reinforcement learning, and FM reinforcement learning. The research team reports that this pipeline improves content consistency, prosodic nature, voice fidelity, visual quality, and consistency.
The model also handles single-pass long-form mixing for up to 3 minutes, robust text editing cases, and a super-resolution vocoder for 48 kHz output.