Paper 2601.15621
Qwen3-TTS Technical Report
01 In brief
Summary
The Qwen3-TTS technical report introduces a family of multilingual, controllable, robust, and streaming text-to-speech models.
Trained on over 5 million hours of speech data across 10 languages, Qwen3-TTS supports 3-second voice cloning, description-based voice design, and fine-grained control.
It uses a dual-track LM architecture with two tokenizers: Qwen-TTS-Tokenizer-25Hz (single-codebook, semantic-rich, with block-wise DiT for streaming) and Qwen-TTS-Tokenizer-12Hz (12.5 Hz multi-codebook, ultra-low-latency, with lightweight causal ConvNet).
The 12Hz variant achieves first-packet latency as low as 97 ms (0.6B) and 101 ms (1.7B).
Training includes three pre-training stages (general, high-quality, long-context) and post-training with DPO, GSPO, and speaker fine-tuning.
Evaluations show state-of-the-art results: lowest WER on Seed-TTS (1.24 for 12Hz-1.7B), superior speaker similarity across all 10 languages vs.
MiniMax and ElevenLabs, 66% error reduction in zh-to-ko cross-lingual generation, and strong instruction following on InstructTTSEval.
The 25Hz variant excels in long speech generation (WER 1.533 zh, 1.571 en).
Models and tokenizers are released under Apache 2.0.
02 From the paper
Abstract
In this report, we present the Qwen3-TTS series, a family of advanced multilingual, controllable, robust, and streaming text-to-speech models. Qwen3-TTS supports state-of-the-art 3-second voice cloning and description-based control, allowing both the creation of entirely novel voices and fine-grained manipulation over the output speech. Trained on over 5 million hours of speech data spanning 10 languages, Qwen3-TTS adopts a dual-track LM architecture for real-time synthesis, coupled with two speech tokenizers: 1) Qwen-TTS-Tokenizer-25Hz is a single-codebook codec emphasizing semantic content, which offers seamlessly integration with Qwen-Audio and enables streaming waveform reconstruction via a block-wise DiT. 2) Qwen-TTS-Tokenizer-12Hz achieves extreme bitrate reduction and ultra-low-latency streaming, enabling immediate first-packet emission ($97\,\mathrm{ms}$) through its 12.5 Hz, 16-layer multi-codebook design and a lightweight causal ConvNet. Extensive experiments indicate state-of-the-art performance across diverse objective and subjective benchmark (e.g., TTS multilingual test set, InstructTTSEval, and our long speech test set). To facilitate community research and development, we release both tokenizers and models under the Apache 2.0 license.