Haihua Xu is visiting Speech through ICSI's visitor agreement with Nanyang Technological University in Singapore.

Haihua received his PhD from Shanghai Jiaotong University in China and is now a researcher at Temasek Laboratories at the Nanyang Technological University. There, he works on acoustic modeling, the statistical representation of sounds that is used by automatic speech recognizers. His focus is on detection of spoken terms, which can be words or strings of words. He's also interested in large-vocabulary speech recognition, subspace Gaussian mixture modeling, and system compilation, or the fusion of multiple systems used to recognize speech. His data is mainly drawn from Malaysian-language news broadcasts. At Temasek, he was a key member of a group that came in fourth in the 2013 NIST Keyword Spotting Competition.

Here at ICSI, he will be working on unsupervised and multilingual acoustic model training, in which very limited human-annontated data is used to train the target model. Unsupervised models use either a subspace Gaussian mixture model training recipe or an unsupervised deep neural network pretraining recipe. In the case of multilingual training, source language transcription is needed and a multilingual-based deep neural network training technique is employed.

Haihua enjoys playing table tennis and swimming.