Contributions

Research Findings

1) We have proposed a data measurement scheme for environmental noise mapping in a city-scale area and developed three new signal processing methods for accurately estimating origins of noise sources and noise intensity levels. The data measurement scheme allows noise data collection for large areas and performance of the new estimation methods significantly outperforms state-of-the-art methods verified using experimental studies.

A paper titled “Large region acoustic source mapping using movable arrays” was presented at ICASSP 2015.

2) We have proposed a learning-based approach for direction-of-arrival (DOA) estimation of sound sources in noisy and reverberant environments. Classic signal processing methods often introduce large DOA estimation errors in noisy and reverberant environments. The learning-based approach uses a model (i.e., MLP neural networks) to directly mapping a noisy observation from a microphone array to the DOA of a sound source. Much smaller estimation errors were achieved by the learning-based approach compared to the classic signal processing methods verified using experiments and real-data tests.

A paper titled “A learning-based approach to direction-of-arrival estimation in noisy and reverberant environments” was presented ICASSP 2015.

3) We proposed a learning-based approach to estimate the reverberation time of enclosures using deep neural networks (DNN) which significantly outperforms state-of-the-art methods.

A paper titled “Learning to estimate reverberation time in noisy and reverberant rooms” was presented INTERSPEECH 2015.

4) We further improved the results of deep neural networks (DNN) based speech de-reverberation for the REVERB Challenge 2014.

A paper titled “Speech dereverberation for enhancement and recognition using deep neural networks and feature adaptation” was submitted to the EURASIP Journal on Advances in Signal Processing, Special Issue on “Silencing the echoes” for possible publication.

5) We proposed a new expectation-maximization eigenvector clustering approach for direction of arrival estimation of multiple speech sources which significantly outperforms state-of-the-art methods in noisy and reverberant environments.

A paper titled “An expectation-maximization eigenvector clustering approach to direction of arrival estimation of multiple speech sources” was submitted to ICASSP 2016 for possible publication.

6) We developed an automatic speech recognition (ASR) system which ranked No. 8 out of 26 systems in the 3^rd CHiME Speech Separation and Recognition Challenge (CHiME-3), an official IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) 2015 Challenge task.

A paper titled “Robust speech recognition using beamforming with adaptive microphones gains and multichannel noise reduction” was to be presented at ASRU 2015.

7) We achieved a significant reduction of computational complexity and improvement of localization and power estimation in acoustic noise source mapping. We also achieved significant performance improvement for acoustic noise mapping by a new proposed method which significantly increases the resolution of acoustic noise source localization, the accuracy of power estimation, and the robustness to ambient noise

Two papers titled “Large region acoustic source mapping: a sparse covariance fitting approach using a movable array” and “Large region acoustic source mapping: a generalized sparse constrained deconvolution approach” were submitted to ICASSP 2016 for possible publication.