Rafal Krolikowski

Neuro-rough control of masking thresholds for audio signal enhancement

Abstract The paper addresses the problem of neuro-rough hybridisation applied to digital processing of audio signals. Moreover, the application of some selected soft computing techniques to non-stationary noise reduction is described. Some attention is also put to a discussion of the intelligent decision algorithms performance. The noise reduction algorithm is based on the new perceptual approach exploiting some properties of the human auditory system. Furthermore, the paper introduces the engineered perceptual filter driven by an intelligent controller employing rules generated with the use of a rough set-based algorithm supported by a neural network. The goal of the intelligent controller is to estimate the current statistics of corrupting noise on the basis of the analysis of signals received from telecommunication channel. Thereafter, the noise estimate enables determining the masking threshold levels which allow making the noise inaudible in the audio signals. Since the implemented decision algorithm requires quantised data, thus the Kohonens self-organising maps (SOM) extended by various distance metrics were used as data quantisers. Some results of the experiments in the domain of non-stationary noise reduction in speech are discussed in the paper.

Noise reduction in audio signals based on the perceptual coding approach

A new concept for the reduction of noise affecting audio signals transmitted in telecommunication channels is proposed. This concept is exploiting some features of the human auditory system. A strong subjective effect of noise suppression in noisy audio can be obtained by uplifting the masking thresholds above the estimated level of the noisy components or by reducing this level in such a way that the components be maintained just below the masking thresholds. The foundations of the engineered method together with the appropriate algorithms are described. A discussion on the results of experiments carried out and some conclusions are also included. The main focus is put on the perceptual foundations of the noise reduction method.

A method for spectral transposition of speech signal applicable in profound hearing loss

Experiments that have been carried out by the authors so far point out that locating the essential components of the speech signal in a low frequency band can improve the speech intelligibility of some hearing impaired people. It is caused by the fact that such persons can maintain their hearing ability in this range of frequencies. The experiments were performed with the use of an electronic device operating in real-time with an algorithm for lowering the audio signals frequency. The satisfactory results of these tests encouraged them to run a research project that involves designing a new digital speech processor and elaborating an advanced algorithm for the spectral transposition of speech. The paper presents a new concept of a hearing aid employing the spectral transposition method. It presents the general scheme of the vocoder-based transposition of speech signals and gives a brief description of the new speech processor implementing this method.

Noise Reduction in Telecommunication Channels using Rough Sets and Neural Networks

A new concept of reduction of non-stationary noise affecting audio signals transmitted in telecommunication channels is proposed. This concept exploits some features of the human auditory system as well as some methods originated from soft computing domain; i.e. rough set-based reasoning and neural processing. The foundations of the engineered method and a description of applied decision algorithms are presented. A number of experiments have been prepared; and some of them have already been carried out. A brief discussion of these experiments’ results and some conclusions are also included.

Noise reduction algorithms employing an intelligent inference engine for multimedia applications

Two approaches to noise reduction are presented, namely the spectral subtraction system and the perceptual coding algorithm allowing to diminish audible noise. Both systems are controlled by an intelligent inference engine based on fuzzy logic. An extension of perceptual coding applications was proposed and verified experimentally with regard to noise removal originally present in acoustic signals. The engineered intelligent systems for noise reduction are presented briefly.

Localization of Sound Sources by Means of Recurrent Neural Networks

The issue of localization of sound sources for videoconferencing is discussed in the paper. A new algorithm for estimating speaker locations, based on recurrent neural networks (RNN), is introduced and described. The scheme of experiments carried out in an acoustically adopted chamber, exploiting the engineered method is detailed.

Echo and noise reduction methods for multimedia communication systems

New concepts of echo cancellation and reduction of non-stationary noise affecting audio signals transmitted in telecommunication channels are proposed in the paper. In the both cases, some methods originated form artificial intelligence domain, i.e.: genetic algorithms, neural networks, rough sets are applied. Moreover, in the noise reduction method, some features of the human auditory system are exploited. A number of experiments have been carried out, and a brief discussion on some of them is included.

Noise reduction in acoustic signals using the perceptual coding and intelligent decision systems

A new algorithm of broadband, nonstationary noise reduction was engineered employing the perceptual approach to the removal of noise from acoustic signals. It enables analysis and processing of sound according to characteristics of hearing sense, and employs a decision system to automatically adjust thresholds of masking. The presented algorithm provides an extension of perceptual coding applications, because it makes possible the reduction of noise included in source signals. Although noisy components of a signal may occur randomly, it is possible to estimate the noise distribution on the basis of signal analysis in silence passages of the transmission. Subsequently, the decision should be made as to the masking curves’ parameter settings which will make the noise inaudible. That is the reason why an intelligent approach is used in modeling some interrelations between available noise patterns and the noise affecting consecutive portions of useful signals. The decision system was implemented in various wa...

Automatic identification of sound source direction based on neural networks

Automatic identification of sound sources direction is still an unsolved problem in contemporary teleconferencing systems. Speech signals coming from various directions not only interfere with the target signal but also can obscure it. People have difficulty in understanding speech with background noise, high reverberation, and/or with many concurrent speakers. Source identification system should allow tracking a target speaker automatically without much delay in order to avoid picking up concurrent speakers by the same microphone channel. In literature one can find two approaches to this problem. One of them is a classical approach to this problem based on delay‐summation algorithms, superdirective arrays and adaptive algorithms, and nonlinear frequency domain microphone array beamformers, etc. The second solution to this problem was proposed among others by the authors in their previous studies, namely, they applied Artificial Neural Networks (ANNs) for the purpose of the automatic sound source localization. In this paper a method for automatic detection of sound source was studied. Both standard feed‐forward ANNs and Recurrent Neural Networks were employed for that purpose. Comparison of the results obtained is given. Conclusions are also derived. [Research sponsored by the Committee for Scientific Research, Warsaw, Poland, Grant No. 8 T11D 00218.]