Alibaba's Qwen launches Qwen-Audio-3.0-TTS, a voice model with multi-language and dialect support

Alibaba's Qwen has launched the Qwen-Audio-3.0-TTS voice synthesis model, with two versions: Flash for real-time interaction (about 300ms latency) and Plus for high-quality output. The new model improves fine-grained label control, freestyle instruction following, and multi-language and dialect coverage, making synthetic speech more expressive.

Users can embed tags like [gasp], [giggles], and [angry] directly in text to control tone and emotion. The model also comes with a premium voice library that turns multi-language, dialect, and instruction capabilities into usable voice resources. It upgrades audio output from 24KHz to 48KHz, with each session lasting up to 3 minutes.

Optimized for global use, it supports 16 languages including Chinese, English, Japanese, Korean, and German, plus new ones like Arabic and Vietnamese. It also covers 20 Chinese dialects like Cantonese and Sichuanese. The Plus version has topped the global Artificial Analysis rankings.