Lead the AI effort for the company:- Large language model: Proprietary LLM training, and LLM based productions for enterprise clients (https://github.com/OrionStarAI/Orion)- AI for Robots: Embodied AI and Robotic Arms, Spoken dialogue, Perception and Vision, SLAM, Planning

Led a machine learning R&D team on speech and NLP:- In-vehicle dialogue system to provide natural interaction for drivers- Safety and compliance detection system serving more than 20M transactions per day- Intelligent customer service system serving 1M customer contacts per day- Open-sourced deep learning based speech and NLP platform (https://github.com/didi/delta)

- Deep learning based large scale NLP platform DeepText- Encoder-decoder modeling for text and speech processing- NLP productions: Question & answering system, dialog system, chatbot- Content based and personalized recommendation system- ASR and speech processing

Recurrent Neural Networks Based Pitch Estimation / 2013 - 2014 - Use deep recurrent neural network to learn features for pitch tracking under noisy conditions. - The approaches significantly improve pitch estimation performance and benefit speech separation and automatic speech recognition.Deep Learning for Dereverberation and Denoising / 2012 - 2014 - Utilized deep neural networks to perform spectral feature mapping for dereverberation and denoising. - The system significantly improves speech quality and automatic speech recognition performances in reverberant and noisy environments.Learning Invariant Features in Complex Environments / 2011 - 2012 - Proposed a kernel learning approach to encode invariant features robust to various noises. - The system only needs a small training set but is generalizable to a large variety of unseen noise conditions without the prior information of the noises.Generalization of Classification Based Speech Separation / 2010 - 2011 - Utilized a rethresholding approach to generalize the trained classifiers to unmatched conditions. - Proposed a distribution fitting method to find the optimal thresholds of support vector machines for speech separation under unseen SNR and noise conditions.A Classification Based Approach to Speech Separation / 2009 - 2010 - Proposed a support vector machines based classification approach to estimate the ideal binary mask for speech separations. - The system improves intelligibility measures and leads to substantial advantages to state-of-the-art.

Deep Neural Networks for Emotion Recognition / 2013 - Use deep neural networks to extract emotional information from speech signals. - Use extreme learning machines for sequence classification for emotion recognition. - The algorithm boosts speech emotion recognition results, which are potentially helpful in user experience enhancement, voice search, call center, and etc.