Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Qwen3-TTS Technical Report

Hangrui Hu, Xinfa Zhu, Ting He, Dake Guo, Bin Zhang, Xiong Wang, Zhifang Guo, Ziyue Jiang, Hongkun Hao, Zishan Guo, Xinyu Zhang, Pei Zhang, Baosong Yang, Jin Xu, Jingren Zhou, Junyang Lin

TL;DR

Qwen3-TTS tackles the demand for multilingual, controllable, and streaming text-to-speech by introducing a dual-track autoregressive framework that unifies textual and acoustic tokens. It deploys two tokenizers—25Hz for balanced semantic-acoustic content and 12Hz for ultra-low latency streaming—enabling real-time synthesis and adaptable voice design. Through extensive pretraining on millions of hours and targeted post-training, Qwen3-TTS achieves state-of-the-art results in zero-shot voice cloning, cross-lingual transfer, and controllable speech generation, while maintaining stability over long-form output. The work highlights practical impact for real-time, LLM-integrated audio systems and is released under an open-source Apache 2.0 license to accelerate community research and development.

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.

Qwen3-TTS Technical Report

TL;DR

Qwen3-TTS tackles the demand for multilingual, controllable, and streaming text-to-speech by introducing a dual-track autoregressive framework that unifies textual and acoustic tokens. It deploys two tokenizers—25Hz for balanced semantic-acoustic content and 12Hz for ultra-low latency streaming—enabling real-time synthesis and adaptable voice design. Through extensive pretraining on millions of hours and targeted post-training, Qwen3-TTS achieves state-of-the-art results in zero-shot voice cloning, cross-lingual transfer, and controllable speech generation, while maintaining stability over long-form output. The work highlights practical impact for real-time, LLM-integrated audio systems and is released under an open-source Apache 2.0 license to accelerate community research and development.

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 () 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.
Paper Structure (26 sections, 3 figures, 10 tables)

This paper contains 26 sections, 3 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Qwen-TTS-Tokenizer
  3. Qwen-TTS-Tokenizer-25Hz
  4. Tokenizer
  5. Streaming Detokenizer
  6. Qwen-TTS-Tokenizer-12Hz
  7. Method
  8. Architectures
  9. Qwen3-TTS-25Hz
  10. Qwen3-TTS-12Hz
  11. Training
  12. Features
  13. Efficiency
  14. Experiments
  15. Evaluation of Speech Tokenizer
  16. ...and 11 more sections

Figures (3)

  • Figure 1: Qwen3-TTS is a multilingual, controllable, robust, and streaming text-to-speech model. Based on these features, Qwen3-TTS supports a wide range of tasks, including but not limited to cloning, creating and controlling voice, and easily handling various complex texts.
  • Figure 2: Overview of Qwen-TTS tokenizers.
  • Figure 3: The overview of Qwen3-TTS. Dashed lines represent optional.