Comparison of techniques for environmental sound recognition

This paper presents a method aimed at recognizing environmental sounds for surveillance and security applications. We propose to apply one-class support vector machines (1-SVMs) together with a sophisticated dissimilarity measure in order to address audio classification, and more specifically, sound recognition. We illustrate the performance of this method on an audio database, which consists of 1015 sounds belonging to nine classes. The database used presents high intraclass diversity in temps of signal properties and some kind of interclass similarities. A large discrepancy in the number of items in each class implies nonuniform probability of sound appearances. The method proceeds as follows: first, the use of a set of state-of-the-art audio features is studied. Then, we introduce a set of novel features obtained by combining elementary features. Experiments conducted on a nine-class classification problem show the superiority of this novel sound recognition method. The best recognition accuracy (96.89%) is obtained when combining wavelet-based features, MFCCs, and individual temporal and frequency features. Our 1-SVM-based multiclass classification approach overperforms the conventional hidden Markov model-based system in the experiments conducted, the improvement in the error rate can reach 50%. Besides, we provide empirical results showing that the single-class SVM outperforms a combination of binary SVMs. Additional experiments demonstrate our method is robust to environmental noise.

Using One-Class SVMs and Wavelets for Audio Surveillance

Recent successful applications of convolutional neural networks (CNNs) to audio classification and speech recognition have motivated the search for better input representations for more efficient training. Visual displays of an audio signal, through various time-frequency representations such as spectrograms offer a rich representation of the temporal and spectral structure of the original signal. In this letter, we compare various popular signal processing methods to obtain this representation, such as short-time Fourier transform (STFT) with linear and Mel scales, constant-Q transform (CQT) and continuous Wavelet transform (CWT), and assess their impact on the classification performance of two environmental sound datasets using CNNs. This study supports the hypothesis that time-frequency representations are valuable in learning useful features for sound classification. Moreover, the actual transformation used is shown to impact the classification accuracy, with Mel-scaled STFT outperforming the other discussed methods slightly and baseline MFCC features to a large degree. Additionally, we observe that the optimal window size during transformation is dependent on the characteristics of the audio signal and architecturally, 2D convolution yielded better results in most cases compared to 1D.

Comparison of Time-Frequency Representations for Environmental Sound Classification using Convolutional Neural Networks.

Sound is one of a human beings most important senses. After vision, it is the sense most used to gather information about the environment. Despite this, comparatively little research has been done into the field of sound recognition. The research that has been done mainly centres around the recognition of speech and music. Our auditory environment is made up of many sounds other than speech and music. This sound information can be taped into for the benefit of specific applications such as security systems. Currently, most researchers are ignoring this sound information. This thesis investigates techniques to recognise environmental non-speech sounds and their direction, with the purpose of using these techniques in an autonomous mobile surveillance robot. It also presents advanced methods to improve the accuracy and efficiency of these techniques. Initially, this report presents an extensive literature survey, looking at the few existing techniques for non-speech environmental sound recognition. This survey also, by necessity, investigates existing techniques used for sound recognition in speech and music. It also examines techniques used for direction detection of sounds. The techniques that have been identified are then comprehensively compared to determine the most appropriate techniques for non-speech sound recognition. A comprehensive comparison is performed using non-speech sounds and several runs are performed to ensure accuracy. These techniques are then ranked based on their effectiveness. The best technique is found to be either Continuous Wavelet Transform feature extraction with Dynamic Time Warping or Mel-Frequency Cepstral Coefficients with Dynamic Time Warping. Both of these techniques achieve a 70% recognition rate. Once the best of the existing classification techniques is identified, the problem of uncountable sounds in the environment can be addressed. Unlike speech recognition, non-speech sound recognition requires recognition from a much wider library of sounds. Due to this near-infinite set of example sounds, the characteristics and complexity of non-speech sound recognition techniques increases. To address this problem, a systematic scheme needs to be developed for non-speech sound classification. Several different approaches are examined. Included is a new design for an environmental sound taxonomy based on an environmental sound alphabet. This taxonomy works over three levels and classifies sounds based on their physical characteristics. Its performance is compared with a technique that generates a structured tree automatically. These structured techniques are compared for different data sets and results are analysed. Comparable results are achieved for these techniques with the same data set as previously used. In addition, the results and greater information from these experiments is used to infer some information about the structure of environmental sounds in general. Finally, conclusions are drawn on both sets of techniques and areas of future research stemming from this thesis are explored.

Non-Speech Environmental Sound Classification System for Autonomous Surveillance

Blind Equalization and System Identification

A speech analysis system, including a kurtosis module for processing a coded sound signal to generate kurtosis measure data; a wavelet module for processing the coded sound signal to generate wavelet coefficients; and a classification module for processing the wavelet coefficients and the kurtosis measure data to generate label data representing a classification for the coded sound signal. The sound signal is classified as environmental noise, silence, speech from a single speaker, speech from multiple speakers, speech from a single speaker plus environmental noise, or speech from multiple speakers plus environmental noise. Speech is further classified as voiced or unvoiced.

Speech analysis system

An algorithm for the analysis of speech utilising the time frequency properties of wavelets is introduced. The extracted wavelet coefficients are analysed using two techniques, firstly a covariance matrix is generated to provide information about speaker characteristics. Second, the kurtosis of the wavelet coefficients is used to facilitate the detection of multiple speakers. Preliminary results show that some phonetic information, such as articulation placement and identification of voiced/unvoiced sections, can be extracted from the kurtosis analysis.

Speech perception based algorithm for the separation of overlapping speech signal

Statistical measures commonly used in signal processing applications such as blind equalisation and optimal filter design are often applied over long periods (seconds). This paper introduces a technique for analysing speech in the time domain using high order statistics. Two measures are investigated kurtosis and skew. Plots of skew and kurtosis are analysed for consistencies and inconsistencies across utterances by a single speaker and multiple speakers. Preliminary results presented show that voiced, unvoiced and silence periods are well characterised using these statistical measures even when applied over a short term (milliseconds).

Speech features found in a continuous high order statistical analysis of speech

A continuous wavelet transform, with Morlet wavelets as the basis functions, is used to map speech into the time-frequency domain. Forward and inverse FFT routines are used to implement the wavelet transforms. A coefficient covariance matrix is defined and an Eigenvalue decomposition is used to optimally determine significant wavelet based filters that accurately represent speech and potentially identify different speakers.

A practical approach to real-time application of speaker recognition using wavelets and linear algebra

This paper describes a wavelet-statistics based feature detection system as applied to microcalcification detection. While a number of researches have been conducted towards microcalcification detection using wavelet analysis and auxiliary information, most of this auxiliary information was obtained from within the spatial domain. In this research, a continuous wavelet transform was used to segment features and compute energy maps of these segmented features. The kurtoses of these features were computed in the wavelet domain. This statistical information together with the energy maps forms the inputs to a rule-based classifier. Physiological information from the spatial domain was used to exclude false-positives. The system was tested using a ROI from the LLNL database. The result is one false-positive within the cluster as classified by the radiologist

Michael Christopher Orr

Papers

Speech analysis system

Speech perception based algorithm for the separation of overlapping speech signal

Speech features found in a continuous high order statistical analysis of speech

A practical approach to real-time application of speaker recognition using wavelets and linear algebra

A Novel Wavelet-Statistics Based Feature Detection System for Detecting Microcalcifications