My responsibilities are:- Lead the team to develop NN-based algorithms including, NR, AEC, beamforming, DoA, multi-channel KWS, and more- Work closely with FW team to deploy algorithms on chip- Planning technology roadmap- Recruiting talent

I developed and realized a very low MIPS and memory footprint ASR/command recognition system. More specifically, I completed following works - AM model training, quantizing- Decoding graph construction, which means user can define their own command set (or grammar).- Decoder implementation and porting. I've designed low MIPS/memory fixed-point decoder which is successfully deployed to our company's AI chip.Other responsibilities are:- Keep tracking on latest ASR technology in industry/academic- Voice UI related roadmap

I joined the Lenovo AI Lab and developed trigger system combined with speaker verification using deep learning framework. I also realized acoustic echo cancellation and beamforming techniques to make trigger and cloud-based ASR more robust.

I developed and realized a very low MIPS and memory footprint voice wake up system from scratch. It has successfully applied on several MTK’s chips, including Helio X20 and X25 series. Moreover, it is adopted by Meizu high-end pro6 smartphone. In addition to wake up feature, I have involved in several ML-related projects, such as gesture recognition with ultrasound, pronunciation decoding with dynamic programming, and audio event detection such as baby crying detection. Several working experiences are listed below:- Voice wake up algorithm design (including training and decoding), implementation, optimization with fixed-point, and integration with CM4 and HAL- Train acoustic model with Kaldi, as well as build up HMM/DNN toolchains- Automate and optimize performance testing flow with c++ and bash/perl- Apply ML methods and dimensionality reduction to classify/visualize problems

I developed a low MIPS fixed-point feature extraction: LPC feature with Levinson–Durbin recursion. I also developed a fixed-point single channel noise reduction method.

At School of Electrical and Computer Engineering, Center for Signal and Image Processing, under the supervision of Prof. Chin-Hui LEE