Dr. Tomohiro Nakatani of NTT Communication Science Laboratories

The convolutional beamformer (CBF) is a signal processing technique that jointly performs denoising, dereverberation, and source separation based on the inverse-filtering framework. It can recover clean, close-talk–quality speech from noisy and reverberant speech mixtures, thereby improving both human listening experiences and automatic speech recognition (ASR) performance.

This talk will introduce the CBF framework, including its formal definition, the mechanism that enables joint enhancement, and its maximum-likelihood–based optimization scheme. It will then show that CBF can be decomposed into Multichannel Linear Prediction (MCLP) for dereverberation and beamforming (BF) for denoising and source separation. This decomposition not only makes the framework more interpretable but also enables flexible configurations that deliver highly effective estimation in real-world tasks. The CHiME-8 distant ASR challenge will be presented as a representative use case.

The talk will also cover recent extensions, including blind CBF, which can operate under unknown recording conditions. In particular, integrating blind CBF with neural-network-based speech enhancement can deliver very high speech quality, even with a small number of microphones and under severely adverse conditions.

Biography:

Tomohiro Nakatani is a Senior Distinguished Researcher at the Communication Science Laboratories, NTT, Inc., Japan. He received his B.E., M.E., and Ph.D. degrees from Kyoto University in 1989, 1991, and 2002, respectively. Since joining NTT in 1991, he has focused on advancing audio signal processing technologies, including speech enhancement and robust automatic speech recognition (ASR). Together with his colleagues, Dr. Nakatani has developed several influential techniques: the blind dereverberation method Weighted Prediction Error (WPE), the blind source separation method complex Angular Central Gaussian Mixture Model (cACGMM), and the target speech extraction method SpeakerBeam. He also made pioneering contributions to the mask‑based beamforming framework. His work has achieved top performance in robust ASR evaluations, including the REVERB Challenge (2014) and the CHiME‑1 and CHiME‑3 Challenges (2011, 2015). He has served as a member on the IEEE Signal Processing Society Audio and Acoustic Signal Processing Technical Committee (2009–2014) and the Speech and Language Processing Technical Committee (2016–2021). He was elevated to IEEE Fellow in 2021.

Address:Kyoto, Japan