Research

1. l1-norm sparse Bayesian Learning

There are many ways to derive sparse solutions, like l1-norm regularization. But what is the optimal sparseness? You may do cross-validation, but is it the best way? My work of sparse Bayesian learning is to define optimal sparseness in a Bayesian framework and shows how to infer it from data. From more details, please see my thesis .

2. Speech dereverberation [ paper (NIPS 2007), slides ]

Sound examples for Fig. 3 in the paper. We carried out 10 experiments in total in a reverberant room. In each experiment, a sentence of speech (approximately 1.5 seconds, and the same for all experiments) was played through a loud speaker (NSW2-326-8A, Aura Sound) and recorded by a matched omnidirectional microphone pair (M30MP, Earthworks). The speaker-microphone positions (and thus RIRs) were different in different experiments.

Anechoic chamber measurement:

1
2
3
4
5
6
7
8
9
10
Left
Right
Estimated

*Left: left microphone signal; Right: right microphone signal; Estimated: the estimated source by our method, blind sparse channel identification.

3. Time-difference-of-arrival (TDOA) estimation [ paper (WASPAA 2007), slides, video 1, video 2, video 3]

Videos 1~3 show how well our method, blind sparse-nonnegative (BSN) channel identification, tracks a moving loud speaker. The performance of phase transform (PHAT) is also presented for comparison.