Cepstral Domains Research Articles

Earthquake Early Warning Starting From 3 s of Records on a Single Station With Machine Learning

AbstractWe introduce the Ensemble Earthquake Early Warning System (E3WS), a set of Machine Learning (ML) algorithms designed to detect, locate, and estimate the magnitude of an earthquake starting from 3 s of P‐waves recorded by a single station. The system is made of six Ensemble ML algorithms trained on attributes computed from ground acceleration time series in the temporal, spectral, and cepstral domains. The training set comprises data sets from Peru, Chile, Japan, and the STEAD global data set. E3WS consists of three sequential stages: detection, P‐phase picking, and source characterization. The latter involves magnitude, epicentral distance, depth, and back azimuth estimation. E3WS achieves an overall success rate in the discrimination between earthquakes and noise of 99.9%, with no false positive (noise mis‐classified as earthquakes) and very few false negatives (earthquakes mis‐classified as noise). All false negatives correspond to M ≤ 4.3 earthquakes, which are unlikely to cause any damage. For P‐phase picking, the Mean Absolute Error is 0.14 s, small enough for earthquake early warning purposes. For source characterization, the E3WS estimates are virtually unbiased, have better accuracy for magnitude estimation than existing single‐station algorithms, and slightly better accuracy for earthquake location. By updating estimates every second, the approach gives time‐dependent magnitude estimates that follow the earthquake source time function. E3WS gives faster estimates than present alert systems relying on multiple stations, providing additional valuable seconds for potential protective actions.

Vocal-Feature-Based Classification of Post-Laryngectomy Patients for Rehabilitation Monitoring

This paper deals with the analysis of substitution voices in patients who underwent partial laryngectomy for laryngeal cancer, with the aim of identifying a reliable methodology to provide an objective evaluation of post-intervention phonatory impairment and of the effectiveness of rehabilitation therapies. The investigated data-set includes 85 patients who underwent type I Open Partial Horizontal Laryngectomy (22 subjects), type II OPHL (32 subjects) and type III OPHL (31 subjects). The available vocal material (reading task and sustained vowel) was pre-processed in order to remove non-harmonic frames from the patients’ records using two different algorithms. After this preliminary step, a series of features that belong to time, spectral and cepstral domains were extracted from the selected harmonic frames. Then, two different comparisons were made between the classes OPHL-I vs OPHL-II+III and the classes OPHL-II+III(I< 5) vs OPHL-II+III(I≥ 5), where the index I (Intelligibility) of the auditory perceptual scale INFVo was assessed during a preliminary evaluation. Two different feature-selection techniques, which are based on the comparison among the probability distributions of the extracted features and the classification performance of a logistic regression model, identified the features with the best discrimination capabilities, which are harmonic-to-noise ratio, fundamental frequency, spectral kurtosis, spectral entropy and mel-frequency cepstral coefficients. The best classification accuracy of 96.5% (5-fold cross validation) was obtained in the comparison OPHL-I vs OPHL-II+III using a logistic regression model that was trained using the 5° and 95° percentile of the fundamental frequency and the 95° percentile of the spectral entropy extracted from the reading task.

Speech intelligibility assessment of dysarthria using Fisher vector encoding

Neuromuscular disorders can lead to dysarthria. Dysarthria is a speech disorder that mainly affects the human motor speech system. It often results in reduced speech intelligibility. Speech intelligibility is one of the parameters to assess the severity of dysarthria. The intelligibility of speech decreases as the severity increases. Automatic speech intelligibility assessment could be reliable and cost-effective compared to the conventional methods which need experienced speech pathologists. Automatic intelligibility assessment includes feature extraction and classification. Features extracted from spectral and cepstral domains are investigated for use in intelligibility assessment. Frame level features are extracted from different signal representations in spectral and cepstral domains. Fisher vector encoding and temporal encoding that uses descriptive statistics are applied to convert frame level features into utterance level. These features are fed to PLDA and ANN classifiers for intelligibility assessment. Overall, the performance of Fisher vector encoding is found to be superior compared to temporal encoding. In all features, the performance of ANN is better in assessing the intelligibility levels compared to PLDA as expected. The STFT, as well as harmonics in spectral domain and CQCC in the cepstral domain, performed well for intelligibility assessment of unseen data of TORGO database.

Mood State Detection in Handwritten Tasks Using PCA-mFCBF and Automated Machine Learning.

In this research, we analyse data obtained from sensors when a user handwrites or draws on a tablet to detect whether the user is in a specific mood state. First, we calculated the features based on the temporal, kinematic, statistical, spectral and cepstral domains for the tablet pressure, the horizontal and vertical pen displacements and the azimuth of the pen’s position. Next, we selected features using a principal component analysis (PCA) pipeline, followed by modified fast correlation–based filtering (mFCBF). PCA was used to calculate the orthogonal transformation of the features, and mFCBF was used to select the best PCA features. The EMOTHAW database was used for depression, anxiety and stress scale (DASS) assessment. The process involved the augmentation of the training data by first augmenting the mood states such that all the data were the same size. Then, 80% of the training data was randomly selected, and a small random Gaussian noise was added to the extracted features. Automated machine learning was employed to train and test more than ten plain and ensembled classifiers. For all three moods, we obtained 100% accuracy results when detecting two possible grades of mood severities using this architecture. The results obtained were superior to the results obtained by using state-of-the-art methods, which enabled us to define the three mood states and provide precise information to the clinical psychologist. The accuracy results obtained when detecting these three possible mood states using this architecture were 82.5%, 72.8% and 74.56% for depression, anxiety and stress, respectively.

Analysis of Smartphone Recordings in Time, Frequency, and Cepstral Domains to Classify Parkinson's Disease.

ObjectivesParkinson’s disease (PD) is the second most common neurodegenerative disorder; it affects more than 10 million people worldwide. Detecting PD usually requires a professional assessment by an expert, and investigation of the voice as a biomarker of the disease could be effective in speeding up the diagnostic process.MethodsWe present our methodology in which we distinguish PD patients from healthy controls (HC) using a large sample of 18,210 smartphone recordings. Those recordings were processed by an audio processing technique to create a final dataset of 80,594 instances and 138 features from the time, frequency, and cepstral domains. This dataset was preprocessed and normalized to create baseline machine-learning models using four classifiers, namely, linear support vector machine, K-nearest neighbor, random forest, and extreme gradient boosting (XGBoost). We divided our dataset into training and held-out test sets. Then we used stratified 5-fold cross-validation and four performance measures: accuracy, sensitivity, specificity, and F1-score to assess the performance of the models. We applied two feature selection methods, analysis of variance (ANOVA) and least absolute shrinkage and selection operator (LASSO), to reduce the dimensionality of the dataset by selecting the best subset of features that maximizes the performance of the classifiers.ResultsLASSO outperformed ANOVA with almost the same number of features. With 33 features, XGBoost achieved a maximum accuracy of 95.31% on training data, and 95.78% by predicting unseen data.ConclusionsDeveloping a smartphone-based system that implements machine-learning techniques is an effective way to diagnose PD using the voice as a biomarker.

High Security and Capacity of Image Steganography for Hiding Human Speech Based on Spatial and Cepstral Domains

A new technique of hiding a speech signal clip inside a digital color image is proposed in this paper to improve steganography security and loading capacity. The suggested technique of image steganography is achieved using both spatial and cepstral domains, where the Mel-frequency cepstral coefficients (MFCCs) are adopted, as very efficient features of the speech signal. The presented technique in this paper contributes to improving the image steganography features through two approaches. First is to support the hiding capacity by the usage of the extracted MFCCs features and pitches extracted from the speech signal and embed them inside the cover color image rather than directly hiding the whole samples of the digitized speech signal. Second is to improve the data security by hiding the secret data (MFCCs features) anywhere in the host image rather than directly using the least significant bits substitution of the cover image. At the recovering side, the proposed approach recovers these hidden features and using them to reconstruct the speech waveform again by inverting the steps of MFCCs extraction to recover an approximated vocal tract response and combine it with recovered pitch based excitation signal. The results show a peak signal to noise ratio of 52.4 dB of the stego-image, which reflect a very good quality and a reduction ratio of embedded data to about (6%–25%). In addition, the results show a speech reconstruction degree of about 94.24% correlation with the original speech signal.

Automatic Multichannel Volcano-Seismic Classification Using Machine Learning and EMD

This article proposes the design of an automatic classifier using the empirical mode decomposition (EMD) along with machine learning techniques for identifying the five most important types of events of the Ubinas volcano, the most active volcano in Peru. The proposed method uses attributes from temporal, spectral, and cepstral domains, extracted from the EMD of the signals, as well as a set of preprocessing and instrument correction techniques. Due to the fact that multichannel sensors are currently being installed in seismic networks worldwide, the proposed approach uses a multichannel sensor to perform the classification, contrary to the usual approach of the literature of using a single channel. The presented method is scalable to use data from multiple stations with one or more channels. The principal component analysis method is applied to reduce the dimensionality of the feature vector and the supervised classification is carried out by means of several machine learning algorithms, the support vector machine providing the best results. The presented investigation was tested with a large database that has a considerable number of explosion events, measured at the Ubinas volcano, located in Arequipa, Peru. The proposed classification system achieved a success rate of more than 90%.

Automatic fish sounds classification.

The work presented in this paper focuses on the use of acoustic systems for passive acoustic monitoring of ocean vitality for fish populations. Specifically, it focuses on the use of acoustic systems for passive acoustic monitoring of ocean vitality for fish populations. To this end, various indicators can be used to monitor marine areas such as both the geographical and temporal evolution of fish populations. A discriminative model is built using supervised machine learning (random-forest and support-vector machines). Each acquisition is represented in a feature space, in which the patterns belonging to different semantic classes are as separable as possible. The set of features proposed for describing the acquisitions come from an extensive state of the art in various domains in which classification of acoustic signals is performed, including speech, music, and environmental acoustics. Furthermore, this study proposes to extract features from three representations of the data (time, frequency, and cepstral domains). The proposed classification scheme is tested on real fish sounds recorded on several areas, and achieves 96.9% correct classification compared to 72.5% when using reference state of the art features as descriptors. The classification scheme is also validated on continuous underwater recordings, thereby illustrating that it can be used to both detect and classify fish sounds in operational scenarios.

Voice assessments for detecting patients with neurological diseases using PCA and NPCA

In this study, we wanted to discriminate between 30 patients who suffer from Parkinson’s disease (PD) and 20 patients with other neurological diseases (ND). All participants were asked to pronounce sustained vowel /a/ hold as long as possible at comfortable level. The analyses were done on these voice samples. Firstly, an initial feature vector extracted from time, frequency and cepstral domains. Then we used principal component analysis (PCA) and nonlinear PCA (NPCA). These techniques reduce the number of parameters and select the most effective ones to be used for classification. Support vector machine and k-nearest neighbor with different kernels was used for classification. We obtained accuracy up to 88% for discrimination between PD patients ND patients using KNN with k equal to three and five.

Voice assessments for detecting patients with Parkinson’s diseases using PCA and NPCA

In this study, we wanted to discriminate between two groups of people. The database used in this study contains 20 patients with Parkinson's disease and 20 healthy people. Three types of sustained vowels (/a/, /o/ and /u/) were recorded from each participant and then the analyses were done on these voice samples. Firstly, an initial feature vector extracted from time, frequency and cepstral domains. Then we used linear and nonlinear feature extraction techniques, principal component analysis (PCA), and nonlinear PCA. These techniques reduce the number of parameters and choose the most effective acoustic features used for classification. Support vector machine with its different kernel was used for classification. We obtained an accuracy up to 87.50 % for discrimination between PD patients and healthy people.

Processing of linear prediction residual in spectral and cepstral domains for speaker information

In this work the linear prediction (LP) residual is processed in spectral and cepstral domains to model the speaker-specific excitation information. In the spectral domain, the excitation energy information is modeled from subband energies (SBE). The excitation periodicity information is modeled by power differences of spectrum in subband (PDSS) measure. This work carries some refinements in the existing methods of extracting SBE and PDSS by exploiting the nature of the excitation spectrum. The SBE and PDSS values are computed from mel warped residual subband spectrum and called as residual mel subband energies (R-MSE) and mel power differences of subband spectra (M-PDSS), respectively. The different speaker recognition studies performed using NIST-99 and NIST-03 databases demonstrate that R-MSE and M-PDSS features represent good speaker information. It is also demonstrated that the excitation energy information can be better modeled in the cepstral domain by residual mel frequency cepstral coefficients (R-MFCC). Furhter, the evidences provided by M-PDSS and R-MFCC features are different and combine well and provides improved recognition performance. The combined evidence from M-PDSS and R-MFCC together with the vocal tract information further improves the performance. Finally, a comparative study on processing the LP residual in temporal, spectral and cepstral domains demonstrates that with a small compromise with the recognition performance, processing LP residual in spectral and cepstral domains provide compact and effective way of representing the excitation information, as compared to temporal processing.

Automated acoustic detection and classification of small aircraft

Automated acoustic detection and classification of low flying aircraft can be used for the prevention of illicit operations. In this presentation, we review past techniques and a new acoustic classification and detection technique. The sound detection algorithm is based on locking in narrow frequency components in the spectra of the recorded aircraft sound. The classification algorithm uses tonal components in the spectral and cepstral domains using discrete time window peak picking. A set of chosen classifiers include: fundamental (pitch) frequency extracted by the cepstral analysis, frequency of tonal components with the maximal amplitude, power spectral estimate in various frequency bands, and number of peaks in predefined frequency and frequency windows. Various small single engine aircraft, ultralights, and helicopters are acoustically detected and classified as they approach the sensor. The algorithm enables simultaneous target detection, reduces noise sensitivity, and minimizes classifier feature space, while maintaining good classification separation. The acoustic classification algorithm was incorporated into the Acoustic Aircraft Detection system developed by Stevens. This combination allowed automated Doppler correction of the harmonic lines, the optimal part of the aircraft signal selection, and establishing classification merits based on aircraft track. [This work was funded by DHS S&amp;T.]

Hidden-Markov-model-based voice activity detector with high speech detection rate for speech enhancement

A new voice activity detection (VAD) algorithm with soft decision output in Mel-frequency domain is developed based on hidden Markov model (HMM) and is incorporated in an HMM-based speech enhancement system. The proposed VAD uses a two-state ergodic HMM representing speech presence and speech absence. The states are constructed from noisy speech and noise HMMs used in the speech enhancement system. This composite model provides a robust detection of speech segments in the presence of noise and obviates the need for extra modeling in HMM-based speech enhancement applications. As the main purpose of the proposed VAD is to detect speech segments accurately, a hang-over mechanism is proposed and is applied on the output of the VAD to improve the speech detection rate. The VAD is integrated in the HMM-based speech enhancement system in Mel-frequency spectral (MFS) and cepstral (MFC) domains. The performance of the proposed VAD, the effectiveness of the hang-over mechanism and the performance of the VAD-integrated speech enhancement system are evaluated on four noise types at different SNR levels. The experimental results confirm the superiority of the proposed VAD compared to the reference methods particularly for speech detection rate at the dominant noisy conditions.

Comparative Evaluation of Single‐Channel MMSE‐Based Noise Reduction Schemes for Speech Recognition

One of the big challenges in the field of Automatic Speech Recognition (ASR) consists in developing suitable solutions able to work properly also in adverse acoustic conditions, like in presence of additive noise and/or in reverberant rooms. Recently a certain attention has been paid to deeply integrate the noise suppressor in the feature extraction pipeline. In this paper, different single‐channel MMSE‐based noise reduction schemes have been implemented both in the frequency and cepstral domains and the related recognition performances evaluated on the AURORA2 and AURORA4 databases, therefore providing a useful reference for the scientific community.

Identification of Autoregressive Systems in Noise Based on a Ramp-Cepstrum Model

In this paper, a new approach for the identification of a minimum-phase autoregressive (AR) system in the presence of a heavy noise is presented. First, a model, valid for both white noise and periodic impulse-train excitations, for the ramp-cepstrum (RC) of the one-sided autocorrelation function of an AR signal is proposed. A residue-based least-squares optimization technique is then employed in conjunction with the RC model to estimate the AR parameters from a noisy output, with a guaranteed system stability. The proposed ramp-cepstral model fitting combines the good features of both the correlation and cepstral domains, and thus provides a more accurate estimate of the parameters in a noisy environment. Extensive simulations are carried out on synthetic AR systems of different orders in the presence of white as well as colored noise. Simulation results demonstrate quite a satisfactory identification performance even for a signal-to-noise ratio of -5 dB, a level at which most of the existing methods fail to provide accurate estimation. To illustrate the suitability of the proposed technique in practical applications, a spectral estimation of a human vocal-tract system is carried out using noise-corrupted natural speech signals.

A Vibroacoustic Model of Selected Human Larynx Diseases

With the present development of digital registration and methods for processing speech it is possible to make effective objective acoustic diagnostics for medical purposes. These methods are useful as all pathologies and diseases of the human vocal tract influence the quality of a patient’s speech signal. Diagnostics of the voice organ can be defined as an unambiguous recognition of the current condition of a specific voice source. Such recognition is based on an evaluation of essential acoustic parameters of the speech signal. This requires creating a vibroacoustic model of selected deformations of Polish speech in relation to specific human larynx diseases. An analysis of speech and parameter mapping in 29-dimensional space is reviewed in this study. Speech parameters were extracted in time, frequency and cepstral (quefrency) domains resulting in diagrams that qualified symptoms and conditions of selected human larynx diseases. The paper presents graphically selected human larynx diseases.

Method of model adaptation for noisy speech recognition by transformation between cepstral and linear spectral domains

A method of model adaptation for noisy speech recognition determines the cepstral mean vector and covariance matrix of adapted noisy speech from the cepstral mean vectors and covariance matrices of speech and noise. The cepstral mean vectors of noise and speech are first transferred into the linear spectral domain, respectively. The linear spectral mean vectors of noise and speech are then combined to obtain a linear spectral mean vector of noisy speech. Next, the linear spectral mean vector of noisy speech is transferred from the linear spectral domain into the cepstral domain, so as to determine the cepstral mean vector of adapted noisy speech. Further, the cepstral covariance matrices of speech and noise are multiplied by a first and a second scaling factor, respectively, and the multiplied cepstral covariance matrices are combined together, so as to determine the cepstral covariance matrix of adapted noisy speech.

Robust hybrid pitch detector

A hybrid pitch detector characterised by parallel analysis of the speech signal in temporal, spectral and cepstral domains is proposed. The voiced/unvoiced decision and pitch period evaluation is realised by a logical analysis of the results from three domains. The experimental analysis shows the robustness of the detector for noisy and telephone speech.

On-line machine condition monitoring using artificial neural networks.

Modern predictive maintenance strategies often require continuous monitoring of rotating machines condition. By detecting incipient failures, expensive unscheduled downtime can be eliminated, unnecessary periodic maintenance procedures avoided, and potentially dangerous machine failures prevented. Vibration signatures have long been recognized as an effective indicator of machine condition. By observing vibration wave forms in the time, frequency, and cepstral domains, human experts are able to assess accurately the condition of bearings, gears, and other machine components. The expense and limited availability of human experts make it highly desirable to automate the vibration interpretation process. This has proven difficult; human pattern recognition ability has exceeded the most sophisticated computer programs. This paper describes a study where artificial neural networks were trained to emulate the ability of a human expert to detect defective rolling element bearings from their vibration signatures. A cascade of probabilistic and backpropagation neural networks learned to agree with the human expert on 100% of the bearing vibration samples, where these were drawn from data outside of the training set. [Work supported by the Electrical Power Research Institute.]