Qwen-Audio-3.1-TTS (fun-resource-shanghai.oss-cn-shanghai.aliyuncs.com)

Researchers have introduced Qwen-Audio-3.1-TTS, a cutting-edge text-to-speech (TTS) system that optimizes for speech synthesis quality while addressing latency, controllability, and multilingual support. The system employs a novel 12.5 Hz low-frame-rate speech tokenizer to enhance efficiency during inference and features a five-stage progressive training paradigm that harmonizes language model and feature model optimization. This robust approach enables the model to generate coherent audio across 16 languages, including 20 Chinese dialects, and allows for long-form synthesis of up to three minutes. The significance of Qwen-Audio-3.1-TTS lies in its state-of-the-art performance on various evaluation metrics, topping the Artificial Analysis Text-to-Speech Leaderboard and demonstrating superior capabilities in content consistency, speaker fidelity, and the handling of noisy input. Its innovative controllability allows users to provide natural-language instructions and fine-grained tags for emotional and stylistic adjustments, including non-verbal cues like laughter and breathing. These advancements not only enhance user experience but also position the model as a leading tool for production-level speech synthesis, paving the way for more natural and accessible AI-human interactions in diverse applications.