Qwen3 TTS Flash is a low-latency text-to-speech model developed by Alibaba, designed for real-time multilingual and multi-dialect speech synthesis. It represents the flagship speech generation capability within the Qwen3 series, achieving first-packet latency as low as 97ms. The model is capable of producing natural, expressive speech across 10 major languages, including Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian.
Built on a dual-track streaming architecture, the model utilizes the Qwen3-TTS-Tokenizer-12Hz, a multi-codebook speech encoder that performs efficient acoustic compression while preserving paralinguistic information and environmental features. This architecture avoids the bottlenecks associated with traditional Diffusion Transformer (DiT) models, allowing for high-speed, high-fidelity reconstruction and adaptive prosody that responds to the semantic nuances of the input text.
Beyond standard text-to-speech, the model family supports advanced features such as voice design and voice cloning. Users can generate a customized voice through natural language descriptions or clone a target timbre using as little as three seconds of reference audio. The system also supports various Chinese dialects, including Mandarin, Cantonese, Hokkien, Wu, and Sichuanese, accurately reproducing regional accents and linguistic nuances.