Jozef Juhár

Feature selection for acoustic events detection

The paper deals with the detection of abnormal situations via captured sound processing. Different settings of feature extraction algorithms were realized and evaluated. Chosen feature sets were used for building the effective parametric representation for gun shots and breaking glass. This way two types of high dimensional feature supervectors were created in regard to the best individual settings of each feature extraction algorithm. For improving the recognition rate Minimum Redundancy Maximum Relevance (MRMR) and Joint Mutual Information (JMI) feature selection algorithms were also applied. They were used for the selection of superior features and for the creation of n-dimensional feature supervectors. The investigation of the appropriate dimension of feature supervectors was performed too. The framework for recognition of potentially dangerous acoustic events such as breaking glass and gun shots, based on the MRMR and JMI selected feature supervector through Hidden Markov Models based classification is proposed in the paper.

Comparison of Different Feature Types for Acoustic Event Detection System

With the increasing use of audio sensors in surveillance or monitoring applications, the detection of acoustic event performed in a real condition has emerged as a very important research problem. This paper is focused on the comparison of different feature extraction algorithms which were used for the parametric representation of the foreground and background sounds in a noisy environment. Our aim was to automatically detect shots and sounds of breaking glass in different SNR conditions. The well known feature extraction method like Mel-frequency cepstral coefficients (MFCC) and other effective spectral features such as logarithmic Mel-filter bank coefficients (FBANK) and Mel-filter bank coefficients (MELSPEC) were extracted from an input sound. Hidden Markov model (HMM) based learning technique performs the classification of mentioned sound categories.

Multimodal biometric authentication using speech and hand geometry fusion

The paper presents biometric security system based on fusion of voice print and hand geometry recognition technologies. Speaker recognition works as text independent and is designed to verify a person using a short utterance. GMM method is used for speaker modeling and GMM-UBM classifier is used for process of matching. Hand geometry technology uses 21 extracted features from image of userpsilas hand and Euclidian distance for recognition. Information fusion in the multimodal system is performed at the matching score level, where scores obtained from matchers are combined using different normalization techniques and fusion rules. Multimodal system after fusion achieved 82.78% reduction in equal error rate over the better of the two independent systems.

Classification of heterogeneous text data for robust domain-specific language modeling

The robustness of n-gram language models depends on the quality of text data on which they have been trained. The text corpora collected from various resources such as web pages or electronic documents are characterized by many possible topics. In order to build efficient and robust domain-specific language models, it is necessary to separate domain-oriented segments from the large amount of text data, and the remaining out-of-domain data can be used only for updating of existing in-domain n-gram probability estimates. In this paper, we describe the process of classification of heterogeneous text data into two classes, to the in-domain and out-of-domain data, mainly used for language modeling in the task-oriented speech recognition from judicial domain. The proposed algorithm for text classification is based on detection of theme in short text segments based on the most frequent key phrases. In the next step, each text segment is represented in vector space model as a feature vector with term weighting. For classification of these text segments to the in-domain and out-of domain area, document similarity with automatic thresholding are used. The experimental results of modeling the Slovak language and adaptation to the judicial domain show significant improvement in the model perplexity and increasing the performance of the Slovak transcription and dictation system.

Service Robot SCORPIO with Robust Speech Interface

The SCORPIO is a small-size mini-teleoperator mobile service robot for booby-trap disposal. It can be manually controlled by an operator through a portable briefcase remote control device using joystick, keyboard and buttons. In this paper, the speech interface is described. As an auxiliary function, the remote interface allows a human operator to concentrate sight and/or hands on other operation activities that are more important. The developed speech interface is based on HMM-based acoustic models trained using the SpeechDatE-SK database, a small-vocabulary language model based on fixed connected words, grammar, and the speech recognition setup adapted for low-resource devices. To improve the robustness of the speech interface in an outdoor environment, which is the working area of the SCORPIO service robot, a speech enhancement based on the spectral subtraction method, as well as a unique combination of an iterative approach and a modified LIMA framework, were researched, developed and tested on simulated and real outdoor recordings.

Acoustic Events Detection Using MFCC and MPEG-7 Descriptors

This paper is focused on the acoustic events detection. Particularly two types of acoustic events (gun shot, breaking glass) were investigated. For any detection task the feature extraction methods play very important role. The feature extraction influences the recognition rate, therefore it is most important in any pattern recognition task. In this paper the impact of Mel-Frequency Cepstral Coefficients - MFCC and selected set of MPEG-7 low-level descriptors were examined. The best feature set contained MFCC and selected descriptors such as ASC, ASS, ASF. They were used to represent the sounds of acoustic events and background. We obtained the improvement of the detection rate using the mentioned set of features. In this task GMM classifiers are used to model the sound classes. This paper describes a basic aspect of our work.

Broadcast news audio classification using SVM binary trees

Audio classification is one of the most important task in content-based analysis and can be implemented in many audio applications, such as indexing and retrieving. This paper addresses the problem of broadcast news audio classification, by support vector machine - binary tree (SVM-BT) architecture, into the five classes: pure speech, speech with music, speech with environment sound, pure music and environment sound. One of the most substantial step in creating such classification architecture is selection of an optimal feature set for each binary SVM classifier. Therefore we implement F-score feature selection algorithm, as an effective search algorithm, within a space of characteristic features that is mostly used for speech/non-speech discrimination.

Performance of Basic Spectral Descriptors and MRMR Algorithm to the Detection of Acoustic Events

This paper is focused on the detection of abnormal situations via sound information. As a main feature extraction algorithm, basic spectral low - level descriptors defined in MPEG-7 standard were used. Various settings for spectral descriptors such as Audio Spectrum Envelope, Audio Spectrum Flatness, Audio Spectrum Centroid and Audio Spectrum Spread were used and many experiments were done for finding the limits of using them for the purpose of acoustic event detection in urban environment. For improving the recognition rate we also applied the feature selection algorithm called Minimum Redundancy Maximum Relevance. The proposed framework of recognizing potentially dangerous acoustic events such as breaking glass and gun shots, based on the extraction of basic spectral descriptors through well known Hidden Markov Models based classification is presented here.

Query-by-example retrieval via fast sequential dynamic time warping algorithm

We introduce a novel approach to Query-by-Example (QbE) retrieval, utilizing fundamental principles of posteriorgram-based Spoken Term Detection (STD), in this paper. Proposed approach is a kind of modification of widely used seg-mental variant of dynamic programming algorithm. Our solution represents sequential variant of DTW algorithm, employing one step forward moving strategy. Each DTW search is carried out sequentially, block by block, where each block represents squared input distance matrix, with size equal to the length of retrieved query. We also examine a way how to speed up sequential DTW algorithm without considerable loss in retrieving performance, by implementing linear time-aligned accumulated distance. The increase of detection accuracy is ensured by weighted cumulative distance score parameter. Therefore, we called this approach Weighted Fast Sequential - DTW (WFS-DTW) algorithm. A novel PCA-based silence discriminator is used along with this algorithm. Evaluation of proposed algorithm is carried out on ParDat1 corpus, using Term Weighted Value (TWV).

Recent advances in the statistical modeling of the Slovak language

In this paper we aim to describe recent advances in the statistical modeling of the Slovak language for transcription of dictated, semi-spontaneous and spontaneous conversational speech such as judicial readings, broadcast news TV and radio shows, parliament proceedings, educational talks and lectures, or interactive conversations. During the last months, we have improved the efficiency and robustness of the Slovak language models trained on the electronic and web-based language resources, including better text processing and document classification, class-based and filled pauses modeling, augmenting of n-grams and fast language model adaptation. Experimental results performed on the judicial readings, broadcast news recordings and parliament proceeding show significant decrease of the word error rate for multiple Slovak transcription system configurations of acoustic and language models in presented scenarios.