Mei-Yuh Hwang
Affiliate Professor at EE Department
University of Washington (UW)

An affiliate professor at University of Washington, Mei-Yuh Hwang is the director of Mobvoi AI Lab in Seattle WA, since early 2016. Mei-Yuh received her PhD in Computer Science from Carnegie Mellon University in 1993 and had worked at Microsoft Seattle and China for 18 years, publishing numerous conference and journal papers, and delivering industry products in speech recognition, machine translation, and language understanding. She is an IEEE fellow, who is passionate in bridging the gap between academia and industry.

Mobvoi makes speech-enabled smart IoT devices, from hardware to software, all in-house, and is constantly looking for AI experts. Though a young company that focuses on industry products, we are actively participating in our speech research community with our limited resources. Our publications can be found in Mobvoi publications.

SPHINX-II speech recognition at CMU, 1987-1993

Mei-Yuh was the first to propose Markov state clustering based on decision trees for continuous speech recognition. The idea of shared states (or senones as Mei-Yuh named it in 1992) has been widely adopted since its inception. Although CTC and end-to-end speech recognition is gaining popularity, most commercial speech recognizers continue to use senones, including the popular open-source toolkit Kaldi.

Speech recognition and machine translation at Microsoft and UW, 1994-2012

Sphinx-II was ported to Microsoft on Windows desktop, Office, and Microsoft Speech Server SDK, for the recognition of multiple languages during 1994-2004. From 2004-2008, Mei-Yuh led the DARPA EARS and GALE Mandarin speech recongition projects at University of Washington (UW). Her IEEE paper was a gold reference guide on building a strong Mandarin speech recognizer.

In 2008-2012, She co-built Bing Translator automated training infrastructure, including the design and implementation of map-reduce parallel processing, based on DryadLink. She further designed and implemented Bing Translation Hub for customized vertical-domain translation.

Spoken language understanding for Cortana, 2012-2015

To deliver non-English Cortana without human annotated data, Mei-Yuh designed an adapted translation algorithm which offered both paraphrasing and generalization capabilities with required slot tags. The protoype model was further improved via iterative data augmentation using RNN and newly logged data. The impressive success of Chinese Cortana gained much attention within China, and sparked the development of personal assistants and AI across China.

Mei-Yuh continued to contribute to Microsoft cognitive services:

Mobvoi AI Lab, Seattle, 2016-present


Professional Services

Invited Talks

Google Scholar


Mei-Yuh's personal home page