Qwen-Audio-3.0-TTS (funaudiollm.github.io)

Qwen Audio has unveiled Qwen-Audio-3.0-TTS, a state-of-the-art text-to-speech synthesis system designed for high production standards. This innovative model boasts significant advancements in content consistency, speaker similarity, and prosodic naturalness, while supporting 16 languages and numerous Chinese dialects. A standout feature is its low-frame-rate speech tokenizer, which operates at 12.5 Hz to minimize inference latency, combined with a five-stage progressive training paradigm that optimizes language modeling and feature mapping (LM and FM) simultaneously. Notably, the model allows for extensive control through natural-language instructions and finely detailed inline tags to regulate emotion, style, and even non-verbal cues like laughter or sighing. The significance of Qwen-Audio-3.0-TTS lies in its capacity for robust deployment in diverse environments, capable of generating high-quality audio even from unclear prompts or under noisy conditions. Winning accolades, including the top position on the Artificial Analysis Text-to-Speech Leaderboard, this synthesis model positions itself as a foundational tool in the AI/ML community. Its comprehensive evaluation approach, including zero-shot voice cloning and acoustic robustness, demonstrates its readiness for real-world applications, enhancing accessibility and user engagement across multilingual platforms.