Alibaba's Qwen-Audio-3.0-TTS tops leaderboard with 16-language, 3-minute speech

Alibaba's Qwen team has unveiled Qwen-Audio-3.0-TTS, a production-oriented speech synthesis system that pushes the boundaries of controllability, robustness, and multilingual coverage. The architecture uses a 12.5 Hz low-frame-rate speech tokenizer to cut inference latency significantly, paired with a five-stage progressive training paradigm that jointly optimizes a language model (LM) and a flow-matching model (FM). This design ensures content consistency, speaker similarity, prosodic naturalness, and high audio quality. Users can control output through free-style natural-language instructions or fine-grained inline tags, enabling precise adjustments to pitch, speed, and emphasis without sacrificing fluency.

The system supports 16 languages and 20 Chinese dialect regions, and can generate up to 3 minutes of continuous speech in a single pass—ideal for audiobooks, virtual assistants, and real-time dubbing. Importantly, it remains robust even with noisy, reverberant, or unclear reference speech, making it practical for real-world production environments. Across benchmarks like SEED-TTS-Eval, CV3-Eval, instruction-following tests, and acoustic-robustness evaluations, Qwen-Audio-3.0-TTS achieves state-of-the-art results on most dimensions. It also claims the top spot on the independent Artificial Analysis Text-to-Speech Leaderboard. These results establish this model as a strong foundation for scalable, high-quality speech synthesis in commercial applications.