Zygmunt Ciota

Speaker verification for multimedia application

In this paper the possibilities of modern speaker verification systems have been described. One of the most important tasks is a proper definition of feature vectors. The long-term spectra vector has been chosen as the background for text-independent verification system. The proposed method permits to compare specific features of the current speaker voice with the computer base containing stored vectors of known speakers. As the results, the distances between the vector of unknown person and the vectors of the persons inclined in the base have been obtained. The possibility of hardware realization of some voice processing components has been also discussed.

Audio-Haptic Feedback in Speech Processing

The main goal of the paper is to achieve better human communication and interaction during conversation process by using a supervising of emotional states and improving voice quality. Therefore, the proposed approach should be also very helpful in the case of vocal tract illness for monitoring of treatment process. Since haptic feedback can nowadays operate by using different sense of touch, like kinesthetic, tactile, cutaneous or force feedback, then such feelings can be also helpful directly in medical treatments as the supplementary method.

Voice processing based on predictive model

The design process of a speech predictor is presented. Going over the recent solutions of neural networks, the systolic structure gives the biggest possibility for smart realization of a speech processor. Mixed analog-digital realizations are also discussed. Practical realizations of current mode building blocks ensure the cheapest realization of CMOS integrated circuits.

Olfactory event-related potentials recordings analysis based on modified EEG registration system

An external effect of the smelling process can be measured as olfactory evoked potentials (OEP). The amplitude of the potentials is in the range of single microvolts. Unfortunately, the precision of such potential measures is low, because of existing “body” noises. Additionally, other brain activities and even human emotion variations can distort the result. Therefore, further researches of electrical properties must be continued. Todays measurement methods of olfactory evoked potentials need several repetition cycles to obtain average values of the necessary potential. Because it is necessary to take into account olfactory weariness (attenuation), a repetition frequency is low and as a consequence, a patient survey is long (usually 20-30 min. using existing equipment).

Design and realization of smart speech processors

In this paper we propose new realizations of small and mobile audio processors applied to speech signal analysis. Proposed algorithms give us a possibility to compare feature vectors of the current speaker voice with the base containing stored vectors of known speakers. Application of microprocessors and integrated circuits dedicated to analysis of audio frequency signals, demands very often modification of several algorithms. Proposed algorithms permit for on-line realization of speaker identification and emotion recognition. The possibility of integrated CMOS realization for different components of audio processing has been also discussed.

Implementation of artificial intelligence methods on example of cardiovascular diseases risk stratification

The paper presents implementation of artificial intelligence methods on the example of the risk stratification of cardiovascular diseases. In the research were used four methods: artificial neural networks, Bayesian networks, decision trees and simple classification models. The obtained results give us possibility to evaluate automatically a health state of the patient and can be helpful in further decisions concerning medical treatment. The obtained results of our research can be also helpful in diagnosis procedures for evaluation of the probability of sudden cardiac arrest.

Modeling of the Speech Process Including Anatomical Structure of the Vocal Tract

The most important features of voice processing have been presented. The properties of glottal waves have been extracted using recorded microphone signals of the speech. Therefore, it was necessary to solve the inverse problem, of finding the glottis input of the whole vocal tract, having the resulting output waves of the speech process. The frequency parameters of glottal waves have been extracted using a vocal tract model. The autocorrelation and cepstrum methods are also helpful in such extraction. The results are important not only for speaker identification and emotion recognition, but can also be helpful for glottis malfunction diagnosis.

Realization of Switched-Current Integrated Circuit with Application of Modified Design-Path

This paper presents a full design-path used to realize a switched-current (SI) ASIC. Novel modifications concerning topology and layout generation are described. Results of successful tests and measurements of manufactured SI ASIC samples are presented.

Compression of wideband sound for multimedia application

Two methods of sound compression: MPEG and Ogg-Vorbis have been discussed and compared. The objective criteria has been used during test procedure. Mixed hardware-software approach can improve the speed and also the quality of the compression. Audible, noise cancellation in high quality speech processing has been also presented. Furthermore, speech coding system has to have real-time capabilities, so the hardware-software co-design permits to achieve low cost and high-speed performances.

Design of A/D converters using current-mode standard cells

In this paper very simple and effective design method and practical realization of current-mode A/D converter (ADC) has been presented. The computer simulations of the proposed structures are very satisfying. The CMOS prototype is very small and consists of four identical cells. The proposed method permits one to design low-voltage, low power and high frequency converters working in a wide range of analog signals.