Prof. Sharon Gannot, Head of the Speech and Signal Processing Laboratory at BIU’s Faculty of Engineering, uses parameter estimation and statistical signal processing for a variety of applications in the field of speech processing.
Gannot’s particular areas of interest are single microphone and microphone array speech enhancement, acoustic echo cancellation, dereverberation, speaker separation and speaker localization. His research team develops algorithms for speech processing utilizing mathematical tools such as robust beamforming, subspace methods, and optimal recursive filtering – the Kalman filter, the extended Kalman, the unscented Kalman and the particle filter.
Recently, Gannot’s team is applying novel statistical inference tools, in particular diffusion maps and non-local filters, to the problem of transient noise reduction and localization.
In this task, Gannot’s team aims to reduce noise levels while imposing the lowest possible distortion on the desired signal received by a single microphone. These algorithms are applicable in many scenarios, where only one microphone can be installed, such as regular cellular phones. Using time domain or frequency domain processing, they are able to reduce slowly time-varying noise signals by either applying the Kalman filter or by modeling the speech signal in the frequency domain utilizing a training stage. Recently, diffusion maps, a modern data inference method, have been successfully applied to the problem of transient noise reduction.
Using spatial information together with time-frequency info, Gannot and his group are able to reduce severe noise levels and extract the desired speaker from multiple speakers – a challenge known as the “cocktail party problem.” A robust beamformer, capable of analyzing the spatial information, was designed and applied in actual acoustic scenarios.
Microphone arrays are also used in dereverberation tasks. In rooms with high levels of reverberation, the original signal can be restored by modeling the room acoustics and by utilizing subspace methods. This application is particularly relevant for settings with several microphones, such as cars, conference calls, and hearing aids. Localization problems are another important application. A speaker (or multiple speakers) can be localized and tracked by applying carefully designed optimal recursive estimation procedures such as the unscented Kalman filter and the particle filter.
For environments with large numbers of microphones, Gannot’s group has derived a distributed and efficient version of the beamforming method that is capable of improving upon the results obtained by regular non-distributed spatial filters. This method is applicable to scenarios with very large number of microphones with unknown positions/locations.
Gannot and his research team are also developing efficient and easily adaptable methods for echo cancellation utilizing statistical room acoustics. Acoustic echo cancellers are known to be sensitive to speaker movements and to double-talk occurrences, especially when the reverberation level is high. In their approach, the acoustic echo path is estimated in two parts. The first part of the echo path is estimated by a regular adaptive filter algorithm (e.g. the least mean squares – LMS), while the contribution of the second is cancelled by application of a spectral subtraction type algorithm.
A new type of acoustic sensor called “Microflown” is being used in Gannot’s lab. This innovative technology enables them to measure both pressure and velocity of the sound field by only one sensor, not larger than a single microphone, Gannot’s group is using the Microflown to develop an algorithm for single sensor localization and speech enhancement.
The unique design of Gannot’s acoustic lab allows for a controllable reverberation time in a large chamber, acoustically isolated from the environment. The lab is fully equipped with state-of-the-art equipment, including 24 microphones, 16 loudspeakers, array installations and sophisticated measurement utilities. This lab is used for the experimental study of most algorithms.