Xu Tan

Research

Tan's research covers language models, multimodality and generative AI for audio and video. In language pretraining, Tan co-authored MASS in 2019, a masked sequence-to-sequence method for language generation, and MPNet in 2020, which combines masked and permuted pretraining for language understanding, both with Kaitao Song, Tao Qin and colleagues.¹⁷⁸

Tan's speech work includes FastSpeech in 2019 and FastSpeech 2 in 2020 for fast, controllable text-to-speech, and the NaturalSpeech series, whose first paper Tan first-authored in 2022. Tan also co-authored HuggingGPT in 2023, an agent framework that uses a language model to plan tasks and call models from Hugging Face, and lists Muzic, a Microsoft project on music understanding and generation, among the systems Tan designed.⁵⁶⁴⁹³¹

Career

At Microsoft Research Asia, Tan was a Principal Research Manager in the Machine Learning Group, and technologies from Tan's work were deployed in Microsoft products including Microsoft Translator, Azure text-to-speech and speech recognition, and Bing. As Research VP of Multimodality at Moonshot AI, Tan has worked on Kimi-Video, Kimi-TTS and Kimi-Audio, whose 2025 technical report lists Tan among its authors.¹¹⁰

Work and contributions

• NaturalSpeech Published research

  First author of NaturalSpeech: End-to-End Text to Speech Synthesis with Human-Level Quality (2022), which uses a variational autoencoder for end-to-end text-to-waveform generation.⁴¹

• FastSpeech and FastSpeech 2 Published research

  Third author, after Yi Ren and Yangjun Ruan, of FastSpeech (2019), and co-author of FastSpeech 2 (2020); Tan's homepage lists both among the audio models Tan designed.⁵⁶¹

• Muzic Open-source code

  Muzic, a project for music understanding and generation started by researchers at Microsoft Research Asia, is published under the MIT license; its README lists Tan as a co-author of several of its papers.³¹

Timeline

1. May 7, 2019Co-authored MASS, a masked sequence-to-sequence pretraining method for language generation.⁷
2. May 22, 2019Co-authored FastSpeech, a fast and controllable text-to-speech model.⁵
3. Apr 20, 2020Co-authored MPNet, combining masked and permuted pretraining for language understanding.⁸
4. Jun 8, 2020Co-authored FastSpeech 2.⁶
5. May 9, 2022Released NaturalSpeech, an end-to-end text-to-speech system, as first author.⁴
6. Mar 30, 2023Co-authored HuggingGPT, which uses a language model to coordinate models from Hugging Face.⁹
7. Apr 25, 2025Listed among the authors of the Kimi-Audio technical report from Moonshot AI.¹⁰

Other documented roles

Roles the sources document that are not confirmed as current.

• Principal Research Manager, Microsoft Research Asia (Machine Learning Group)¹²

Where they work

• Moonshot AI 2 people in the wiki

Related people

• Yuxin WuAlso at Moonshot AI
• Wei-Ning HsuSpeech and audio · Multimodal models
• Alexandre DéfossezSpeech and audio · Image, video, and 3D generation
• Amit JainImage, video, and 3D generation · Multimodal models
• Neil ZeghidourSpeech and audio · Image, video, and 3D generation
• Saining XieImage, video, and 3D generation · Multimodal models
• Nataniel RuizImage, video, and 3D generation · Multimodal models
• Sanyuan ChenSpeech and audio · Image, video, and 3D generation
• Yi RenSpeech and audio · Image, video, and 3D generation
• Alexei BaevskiSpeech and audio · Multimodal models