Speaker Characterization

This study aimed to analyze the impact of the amount of data on the discriminatory performance of acoustic-phonetic parameters, some of which are frequently assessed in forensic speaker comparisons. Parameters from three distinct phonetic domains were considered, namely, spectral, melodic, and temporal, which were assessed separately within the same phonetic domain and in combination. The speech material consisted of spontaneous telephone conversations between two subjects. During the recording sessions, the participants were placed in different rooms, not directly seeing, hearing, or interacting with each other. The speakers were encouraged to start a conversation using a mobile phone while being simultaneously recorded. All recordings were carried out with a high resolution (44.1 kHz and 16-bit). Data segmentation and transcription were performed in the Praat software [1]. The participants were 20 male subjects, Brazilian Portuguese speakers from the same dialectal area. Their age ranged from 19 to 35 years, with a mean of 26.4 years. Although the subjects (10 identical twin pairs) were recruited from a twin research project, cf. [2, 3, 4], the focus here was comparisons among all speakers (i.e., 190 inter-speaker comparisons) rather than on individual twin pairs. Two metrics of discriminatory performance were examined through the R software [5] as a function of the comparisons among all speakers in the study using the script fvclrr [6]: Log-likelihoodratio-cost (Cllr) and Equal Error Rate (EER) values. The Cllr metric is an empirical estimate of the precision of likelihood ratios. The EER metric captures the point where the false reject rate (type I error) and false accept rate (type II error) are equal and is used to describe the overall accuracy of a system. Lower Cllr and EER values are compatible with better discriminatory performance, whereas higher values suggest the opposite trend. A cross-validation procedure was adopted for the calculation of likelihood ratios, where multiple pairwise comparisons were performed across individuals in which the background sample consisted of data from all speakers, except those being directly compared. To assess the impact of the amount of data on discriminatory performance, a very straightforward approach was adopted. Based on a larger data set extracted from recordings of about 2.50 min per speaker, random data points were additively selected for the analyses. The minimum number of data points, i.e., acoustic measurements, selected per speaker was set at 2 to allow the minimum comparison of intra-speaker variability. Thereafter, more randomly selected data points were progressively added to the tests, two points at a time, and new Cllr and EER values were computed for the new resampled data set. For the present study, the maximum number of data points (acoustic measurements) was set at 30. Given the nature of the speech material assessed, a discrepancy in the number of samples produced per subject was observed. Because of that, a random downsampling procedure was repeated 200 times to minimize the selection bias using the R package Recipes [7]. Cllr and EER median values were reported after performing tests with the randomly selected data while performing a fusion and calibration of different estimates based on a logistic regression technique, cf. [8]. Melodic and temporal parameters were extracted from speech chunks with an average (mean and median) temporal window of 3 s, corresponding to inter-pause intervals in most cases. Spectral parameters were extracted from the midpoints of /a/ vowels in stressed and unstressed positions. The monophthongs displayed a mean and a median duration of 67 ms and 84 ms, respectively. After a manual segmentation, all parameters were extracted automatically using a Praat script [9]. Four models were compared. Model 1 (M1) comprises the combination of melodic parameters (f0 median and f0 base value). Model 2 (M2) comprises the combination of temporal parameters (speech rate and articulation rate). Model 3 (M3) consists of the combination of spectral parameters (F3 and F4). Model 4 (M4) considered the combination of all the acoustic-phonetic parameters.

Objectives. To assess the speaker-discriminatory potential of a set of fundamental frequency estimates in intraidentical twin pair comparisons and cross-pair comparisons (i.e., among all speakers). Participants. A total of 20 Brazilian Portuguese speakers of the same dialect, namely 10 male identical twin pairs aged between 19 and 35, were recruited. Method. The participants were recorded directly through professional microphones while taking part in a spontaneous dialogue over mobile phones. Acoustic measurements were performed in connected speech samples, and in lengthened vowels, at least 160 ms long produced during spontaneous speech. Results. f 0 baseline, central tendency, and extreme values were found mostly discriminatory in intra-twin pair and cross-pair comparisons. These were also the estimates displaying the largest effect sizes. Overall, only three identical twins were found statistically different regarding their f 0 patterns in connected speech, but not for lengthened vowel-based f 0 metrics. Estimates of f 0 variation and modulation were found the least discriminatory across speakers, which may signal the control of speaking style and dialect on dynamic patterns of f 0. Concerning system performance, the base value of f 0 (f 0 baseline) was found the most reliable metric, displaying the lowest equal error rate (EER). Conclusions. The outcomes suggest that, although identical twins were very closely related regarding their f 0 patterns, some pairs could still be differentiated acoustically, only in connected speech. Such findings reinforce the relevance of analyzing long-term f 0 metrics for speaker comparison purposes, with particular consideration to f 0 baseline. Furthermore, f 0 differences across subjects were suggested as more expressive in connected speech than in lengthened vowels.T a g g e d E n d T a g g e d P Keywords: Speaking fundamental frequency−Acoustic phonetics−Forensic phonetics−Identical twins.T a g g e d E n d

In forensic voice comparison, it is strongly recommended to follow the Bayesian paradigm to present a forensic evidence to the court. In this paradigm, the strength of the forensic evidence is summarized by a likelihood ratio (LR). But in the real world, to base only on the LR without looking to its degree of reliability does not allow experts to have a good judgement. This work is mainly motivated by the need to quantify this reliability. In this concept, we think that the presence of speaker specific information and its homogeneity between the two signals to compare should be evaluated. This paper is dedicated to the latter, the homogeneity. We propose an information theory based homogeneity measure which determines whether a voice comparison is feasible or not.

This study investigated the production of six Cantonese tones by heritage language (HL) children in Vancouver, Canada. Twenty-five Cantonese heritage speakers (HSs) aged between 2;1 and 6;0 participated in the production experiment. Data collected from children in Hong Kong (homeland) at the same ages and native speakers aged between 15;09 and 16;07 were included for analysis. Results showed that HSs have not fully acquired all six lexical tones, in particular the tonal contrasts, by 6;0. HL children's overall pattern of the pitch contours was similar to that of homeland children, but with smaller tonal distinctions between similar tone pairs relative to reference speakers. An ongoing tone merging phenomenon appeared among the observed HL children. More interestingly, a "highlow" template, resembling the pitch pattern of English trochaic words, was found in the production of two HL children respectively at age 3;1 and 5;0. These findings are consistent with the idea that language variation exists in the acquisition of HL. Furthermore, the onset of schooling leads to the shift in language dominance from the HL to the majority language, which influences the production of HL.

Automatic height and age prediction of a speaker has a wide variety of applications in speaker profiling, forensics etc. Often in such applications only a few seconds of speech data is available to reliably estimate the speaker parameters. Traditionally, age and height were predicted separately using different estimation algorithms. In this work, we propose a unified DNN architecture to predict both height and age of a speaker for short durations of speech. A novel initialization scheme for the deep neural architecture is introduced, that avoids the requirement for a large training dataset. We evaluate the system on TIMIT dataset where the mean duration of speech segments is around 2.5s. The DNN system is able to improve the age RMSE by at least 0.6 years as compared to a conventional support vector regression system trained on Gaussian Mixture Model mean supervectors. The system achieves an RMSE error of 6.85 and 6.29 cm for male and female height prediction. In case of age estimation, the RMSE errors are 7.60 and 8.63 years for male and female respectively. Analysis of shorter speech segments reveals that even with 1 second speech input the performance degradation is at most 3% compared to the full duration speech files.

Support vector machine (SVM) has been proven as a powerful tool for solving age and gender classification problems. However, SVM is sensitive to noise and outliers. In this paper we propose a new fuzzy SVM based on an assumption that training data points should not be treated equally to avoid the problem of sensitivity to noise and outliers. This can be achieved by assigning a fuzzy membership as a weight to each training data point. A method to calculate fuzzy memberships is also presented. Experiments performed on the aGender corpus for INTERSPEECH 2010 Paralinguistic Challenge show that the proposed fuzzy SVM can improve age and gender classification accuracy.

Support vector machine (SVM) has been proven as a powerful tool for solving age and gender classification problems. However, SVM is sensitive to noise and outliers. In this paper we propose a new fuzzy SVM based on an assumption that training data points should not be treated equally to avoid the problem of sensitivity to noise and outliers. This can be achieved by assigning a fuzzy membership as a weight to each training data point. A method to calculate fuzzy memberships is also presented. Experiments performed on the aGender corpus for INTERSPEECH 2010 Paralinguistic Challenge show that the proposed fuzzy SVM can improve age and gender classification accuracy.

Support vector machine (SVM) has been proven as a powerful tool for solving age and gender classification problems. However, SVM is sensitive to noise and outliers. In this paper we propose a new fuzzy SVM based on an assumption that training data points should not be treated equally to avoid the problem of sensitivity to noise and outliers. This can be achieved by assigning a fuzzy membership as a weight to each training data point. A method to calculate fuzzy memberships is also presented. Experiments performed on the aGender corpus for INTERSPEECH 2010 Paralinguistic Challenge show that the proposed fuzzy SVM can improve age and gender classification accuracy.

Support vector machine (SVM) has been proven as a powerful tool for solving age and gender classification problems. However, SVM is sensitive to noise and outliers. In this paper we propose a new fuzzy SVM based on an assumption that training data points should not be treated equally to avoid the problem of sensitivity to noise and outliers. This can be achieved by assigning a fuzzy membership as a weight to each training data point. A method to calculate fuzzy memberships is also presented. Experiments performed on the aGender corpus for INTERSPEECH 2010 Paralinguistic Challenge show that the proposed fuzzy SVM can improve age and gender classification accuracy.

Support vector machine (SVM) has been proven as a powerful tool for solving age and gender classification problems. However, SVM is sensitive to noise and outliers. In this paper we propose a new fuzzy SVM based on an assumption that training data points should not be treated equally to avoid the problem of sensitivity to noise and outliers. This can be achieved by assigning a fuzzy membership as a weight to each training data point. A method to calculate fuzzy memberships is also presented. Experiments performed on the aGender corpus for INTERSPEECH 2010 Paralinguistic Challenge show that the proposed fuzzy SVM can improve age and gender classification accuracy.

Support vector machine (SVM) has been proven as a powerful tool for solving age and gender classification problems. However, SVM is sensitive to noise and outliers. In this paper we propose a new fuzzy SVM based on an assumption that training data points should not be treated equally to avoid the problem of sensitivity to noise and outliers. This can be achieved by assigning a fuzzy membership as a weight to each training data point. A method to calculate fuzzy memberships is also presented. Experiments performed on the aGender corpus for INTERSPEECH 2010 Paralinguistic Challenge show that the proposed fuzzy SVM can improve age and gender classification accuracy.

The objective of the proposed work is to analyze and study the use of i-vectors for Anomalous Detection of Sounds (ADS) in Machines. I-vectors, to the best of our knowledge, have not been studied for machine sounds. We will be using the database ToyADMOS for testing both supervised and unsupervised ADS techniques using i-vectors generated from the data.M.S

Support vector machine (SVM) has been proven as a powerful tool for solving age and gender classification problems. However, SVM is sensitive to noise and outliers. In this paper we propose a new fuzzy SVM based on an assumption that training data points should not be treated equally to avoid the problem of sensitivity to noise and outliers. This can be achieved by assigning a fuzzy membership as a weight to each training data point. A method to calculate fuzzy memberships is also presented. Experiments performed on the aGender corpus for INTERSPEECH 2010 Paralinguistic Challenge show that the proposed fuzzy SVM can improve age and gender classification accuracy.

In the provision of linguistic evidence as one of the foci in Forensic Linguistics, Forensic Speaker Verification (FSV) includes an analysis of speech recordings to verify the voice of a criminal. As an inquiry into the validity of the available FSV, we present the analysis on Indonesian FSV system. The system consists of pairing, tagging, acoustic features extraction, and statistical analysis. There is a claim that the system meets the demand for presenting legal evidence in Indonesian court. In the system, one of the acoustic features extracted from the speech data is fundamental frequency (F0). Then, the paper aims at reviewing the method in Indonesian FSV system in terms of fundamental frequency (F0) used as the discriminatory potential. The results show that F0 has not represented an adequate interpretation of the linguistic evidence in our experimental data. It leads us to suggest that more experimental studies are required to scrutinize F0 in the system.

This paper demonstrates the potential of the sub-band parametric cepstral distance (PCD) formulated by Clermont and Mokhtari (1994), as an alternative to formants in acoustic phonetic research. As a cepstrum-based measure, the PCD is automatically and reliably extracted from the speech signal. By contrast, formants are time -consuming and often difficult to estimate, a well-known bottleneck for studies based on large-scale datasets. The PCD measure gives flexibility in selecting the frequency limits of any sub-band of interest within the available full band. We suggest that, if sub-band selection were guided by the acoustic–phonetic theory of speech production, PCD analysis could facilitate phonetically meaningful cepstral comparisons without relying directly on formants. We evaluate this idea by exploiting the PCD properties in the context of forensic voice comparison as an application example. The cepstral data were obtained from the vowels uttered by 306 male Japanese speakers. Similar patterns of results were observed using formants and sub-band PCDs, the latter yielding better performance. This suggests that sub-band PCDs are able to capture the spectral characteristics that we normally quantify through formants, but with better reliability and efficiency. The PCD results reported here are encouraging for other types of acoustic-phonetic studies in which comparisons of spectral characteristics are required.

Purpose Previous studies showed both early and late acquisition of Cantonese tones based on transcription data using different criteria, but very little acoustic data were reported. Our study examined Cantonese tone acquisition using both transcription and acoustic data, illustrating the early and protracted aspects of Cantonese tone acquisition. Method One hundred fifty-nine Cantonese-speaking children aged between 2;1 and 6;0 (years;months) and 10 reference speakers participated in a tone production experiment based on picture naming. Natural production materials with 30 monosyllabic words were transcribed by two native judges. Acoustic measurements included overall tonal dispersion and specific contrasts between similar tone pairs: ratios of average fundamental frequency height for the level tones (T1, T3, T6), magnitude of rise and inflection point for the rising tones (T2, T5), magnitude of fall, H1*–H2*, and harmonic-to-noise ratio for the low tones (T4, T6). Auditory assessme...

Recent advances in the field of speaker recognition have proved to highly outperform algorithms. However this performance degrades when limited data are presented. This paper presents examples on how SVM can improve speaker recognition. The main contribution in this approach is the use of new low-dimensional vectors when training data are limited. We show how different kernels function of Support Vector M achines (SVM) can be used to deal a new approach for speaker recognition system. We achieved remarkable results using TIM IT database

Current Automatic Speaker Recognition (ASR) System has emerged as an important medium of confirmation of identity in many businesses, ecommerce applications, forensics and law enforcement as well. Specialists trained in criminological recognition can play out this undertaking far superior by looking at an arrangement of acoustic, prosodic, and semantic attributes which has been referred to as structured listening. An algorithmbased system has been developed in the recognition of forensic speakers by physics scientists and forensic linguists to reduce the probability of a contextual bias or pre-centric understanding of a reference model with the validity of an unknown audio sample and any suspicious individual. Many researchers are continuing to develop automatic algorithms in signal processing and machine learning so that improving performance can effectively introduce the speaker's identity, where the automatic system performs equally with the human audience. In this paper, I examine the literature about the identification of speakers by machines and humans, emphasizing the key technical speaker pattern emerging for the automatic technology in the last decade. I focus on many aspects of automatic speaker recognition (ASR) systems, including speaker-specific features, speaker models, standard assessment data sets, and performance metrics.

Speech under face cover constitute a case that is increasingly met by forensic speech experts. Wearing face cover mostly happens when an individual strives to conceal his or her identity. Based on the material of face cover and the level of contact with speech production organs, speech production becomes affected by face mask and a part of speech energy gets absorbed in the mask. There has been little research on how speech acoustics is affected by different face masks and how face covers might affect performance of automatic speaker recognition systems. In the present paper, we have collected speech under face mask with the aim of studying the effects of wearing different masks on state-of-the-art text-independent automatic speaker recognition system. The preliminary speaker recognition rates along with mask identification experiments are presented in this paper.

Most of the existing literature on i-vector-based speaker recognition focuses on recognition problems, where i-vectors are extracted from speech recordings of sufficient length. The majority of modeling/recognition techniques therefore simply ignores the fact that the i-vectors are most likely estimated unreliably when short recordings are used for their computation. Only recently, were a number of solutions proposed in the literature to address the problem of duration variability, all treating the i-vector as a random variable whose posterior distribution can be parameterized by the posterior mean and the posterior covariance. In this setting the covariance matrix serves as a measure of uncertainty that is related to the length of the available recording. In contract to these solutions, we address the problem of duration variability through weighted statistics. We demonstrate in the paper how established feature transformation techniques regularly used in the area of speaker recognition, such as PCA or WCCN, can be modified to take duration into account. We evaluate our weighting scheme in the scope of the i-vector challenge organized as part of the Odyssey, Speaker and Language Recognition Workshop 2014 and achieve a minimal DCF of 0.280, which at the time of writing puts our approach in third place among all the participating institutions.

An acoustic-phonetic forensic-voice-comparison system extracted information from the formant trajectories of tokens of Standard Chinese /iau/. When this information was added to a generic automatic forensic-voice-comparison system, which did not itself exploit acoustic-phonetic information, there was a substantial improvement in system validity but a decline in system reliability.

An acoustic-phonetic forensic-voice-comparison system extracted information from the formant trajectories of tokens of Standard Chinese /iau/. When this information was added to a generic automatic forensic-voice-comparison system, which did not itself exploit acoustic-phonetic information, there was a substantial improvement in system validity but a decline in system reliability.

Examples are given of forensic voice comparison with higher level features in real-world cases and research. A pilot experiment relating to estimation of strength of evidence in forensic voice comparison is described which explores the use of higher-level features extracted over a disyllabic word as a whole, rather than over individual monosyllables as conventionally practiced. The trajectories of the first three formants and tonal F0 of the hexaphonic disyllabic Cantonese word daihyat 'first' from controlled but natural non-contemporaneous recordings of 23 male speakers are modeled with polynomials, and multivariate likelihood ratios estimated from their coefficients. Evaluation with the log likelihood ratio cost validity metric C llr shows an optimum performance is obtained, surprisingly, with lower order polynomials, with F2 requiring a cubic fit, and F1 and F3 quadratic. Fusion of F-pattern and tonal F0 results in considerable improvement over the individual features, reducing the C llr to ca. 0.1. The forensic potential of the daihyat data is demonstrated by fusion with three other Cantonese higher-level features: the F-pattern of /i/, short-term F0, and syllabic nasal cepstral spectrum, which reduces the C llr still further to 0.03. Important pros and cons of higher-level features and likelihood ratios are discussed, the latter illustrated with data from Japanese, and three varieties of English in real forensic casework.

This paper proposes a new speaker age estimation method that uses an age-dependent insensitive loss. Most conventional speaker age estimation frameworks ignore the ambiguity of a perceptual speaker age. These “over-sensitive” frameworks can cause critical errors far from the actual age. We propose an age-dependent insensitive loss for speaker age estimation. The key idea of the proposed method is that the age estimator should allow some ambiguity of actual age and this ambiguity should depend on age. The age-dependent insensitivity is learned by 5-MAE (mean absolute error) loss and soft target cross entropy loss in regression and classification problems. Experimental results showed that the proposed method improves the mean absolute error and the ratio of critical error by 5.2% and 5.7% for the regression problem and 9.6% and 31.5% for the classification problem.

We propose two machine learning improvements on the existing architecture of voiceand speakerrecognition software. Where conventional systems extract two kinds of frequency data from voice recordings and use the concatenation as input, we propose two methods to allow the input vectors to interact multiplicatively. The first is a Neural Tensor Network layer under a softmax classifier, and the second is a constrained variant of the Neural Tensor Network with reduced dimensionality. We compare these methods with the current approach, using SVMs on the concatenation of extracted data vectors. Second, we trained a shallow neural network on a Q-learning framework in order to intelligently and dynamically minimize the amount of audio required to make an accurate classification decision. While the neural network architectures failed to improve on the existing SVM model, the Q-learner did learn to dynamically minimize audio sampling while improving on the accuracy of the SVM system. Keywords...

Voice controlled applications can be a great aid to society, especially for physically challenged people. However this requires robustness to all kinds of variations in speech. A spoken language understanding system that learns from interaction with and demonstrations from the user, allows the use of such a system in different settings and for different types of speech, even for deviant or impaired speech, while also allowing the user to choose a phrasing. The user gives a command and enters its intent through an interface, after which the model learns to map the speech directly to the right action. Since the effort of the user should be as low as possible, capsule networks have drawn interest due to potentially needing little training data compared to deeper neural networks. In this paper, we show how capsules can incorporate multitask learning, which often can improve the performance of a model when the task is difficult. The basic capsule network will be expanded with a regularis...

Audio classification has applications in a variety of contexts, such as automatic sound analysis, supervised audio segmentation and in audio information search and retrieval. Extended Baum-Welch (EBW) transformations are most commonly used as a discriminative technique for estimating parameters of Gaussian mixtures, though recently they have been applied in unsupervised audio segmentation. In this paper, we extend the use of these transformations to derive an audio classification algorithm. We find that our method outperforms both the Support Vector Machine (SVM) and Gaussian Mixture Model (GMM) likelihood classification methods.

Speech signals convey important paralinguistic information such as age, gender, body size, language, accent and emotional state of speakers. Automatic identification of speaker traits and states has a wide range of forensic, commercial and medical applications in real-world scenarios. This thesis proposes a novel approach for automatic estimation of four forensically important components of speaker profiling systems, namely speaker age, height, weight and smoking habit estimations, from spontaneous telephone speech signals. In this method, each utterance is modeled using the i-vector framework which is based on the factor analysis on Gaussian Mixture Model (GMM) mean supervectors, and the Non-negative Factor Analysis (NFA) framework which is based on a constrained factor analysis on GMM weight supervectors. Then, Artificial Neural Networks (ANNs) and Least Squares Support Vector Regression (LSSVR) are employed to estimate age, height and weight of speakers from given utterances. Var...

Voice recognition systems are used to distinguish different sorts of voices. However, recognizing a voice is not always successful due to the presence of different parameters. Hence, there is a need to create a set of estimation criteria and a learning process using Artificial Neural Network (ANN). The learning process performed using ANN allows the system to mimic how the brain learns to understand and differentiate among voices. The key to undergo this learning is to specify the free parameters that will be adapted through this process of simulation. Accordingly, this system will store the knowledge processed after performing the back propagation learning and will be able to identify the corresponding voices. The proposed learning allows the user to enter a number of different voices to the system through a microphone.

Speech-based communication is one of the most preferred modes of communication for humans. The human voice contains several important information and clues that help in interpreting the voice message. The gender of the speaker can be accurately guessed by a person based on the received voice of a speaker. The knowledge of the speaker’s gender can be a great aid to design accurate speech recognition systems. GMM based classifier is a popular choice used for gender detection. In this paper, we propose a Tensor-based approach for detecting the gender of a speaker and discuss its implementation details for low resourceful languages. Experiments were conducted using the TIMIT and SHRUTI dataset. An average gender detection accuracy of 91% is recorded. Analysis of the results with the proposed method is presented in this paper.

This paper proposes an automatic smoking habit detection from spontaneous telephone speech signals. In this method, each utterance is modeled using i-vector and non-negative factor analysis (NFA) frameworks, which yield low-dimensional representation of utterances by applying factor analysis on Gaussian mixture model means and weights respectively. Each framework is evaluated using different classification algorithms to detect the smoker speakers. Finally, score-level fusion of the i-vector-based and the NFA-based recognizers is considered to improve the classification accuracy. The proposed method is evaluated on telephone speech signals of speakers whose smoking habits are known drawn from the National Institute of Standards and Technology (NIST) 2008 and 2010 Speaker Recognition Evaluation databases. Experimental results over 1194 utterances show the effectiveness of the proposed approach for the automatic smoking habit detection task.

This paper proposes an automatic smoking habit detection from spontaneous telephone speech signals. In this method, each utterance is modeled using i-vector and non-negative factor analysis (NFA) frameworks, which yield low-dimensional representation of utterances by applying factor analysis on Gaussian mixture model means and weights respectively. Each framework is evaluated using different classification algorithms to detect the smoker speakers. Finally, score-level fusion of the i-vector-based and the NFA-based recognizers is considered to improve the classification accuracy. The proposed method is evaluated on telephone speech signals of speakers whose smoking habits are known drawn from the National Institute of Standards and Technology (NIST) 2008 and 2010 Speaker Recognition Evaluation databases. Experimental results over 1194 utterances show the effectiveness of the proposed approach for the automatic smoking habit detection task.

This project presents an approach to classify speakers on the basis of their age and gender. Short term features and long term features have been extracted from the voice sample of each speaker. These have been used to train Support Vector Machines and SVM Regressor. For training instances the results from these are combined to obtain the age estimate and gender of the speaker. WSNMF has been used to reduce the dimension of the short term feature space. It has performance improvement in terms of accuracy and computation time. Keywords—Age Recognition, LTF, STF, WSNMF, SVM.

This paper focuses on the automatic detection of a person's blood level alcohol based on automatic speech processing approaches. We compare 5 different feature types with different ways of modeling. Experiments are based on the ALC corpus of IS2011 Speaker State Challenge. The classification task is restricted to the detection of a blood alcohol level above 0.5 ‰. Three feature sets are based on spectral observations: MFCCs, PLPs, TRAPS. These are modeled by GMMs. Classification is either done by a Gaussian classifier or by SVMs. In the later case classification is based on GMM-based supervectors, i.e. concatenation of GMM mean vectors. A prosodic system extracts a 292-dimensional feature vector based on a voicedunvoiced decision. A transcription-based system makes use of text transcriptions related to phoneme durations and textual structure. We compare the stand-alone performances of these systems and combine them on score level by logistic regression. The best stand-alone performance is the transcriptionbased system which outperforms the baseline by 4.8 % on the development set. A Combination on score level gave a huge boost when the spectral-based systems were added (73.6 %). This is a relative improvement of 12.7 % to the baseline. On the test-set we achieved an UA of 68.6 % which is a significant improvement of 4.1 % to the baseline system.

Autism Spectrum Disorder (ASD) is on the rise and constantly growing. Earlier identify of ASD with the best outcome will allow someone to be safe and healthy by proper nursing. Humans can hardly estimate the present condition and stage of ASD by measuring primary symptoms. Therefore, it is being necessary to develop a method that will provide the best outcome and measurement of ASD. This paper aims to show several measurements that implemented in several classifiers. Among them, Support Vector Machine (SVM) provides the best result and under SVM, there are also some kernels to perform. Among them, the Gaussian Radial Kernel gives the best result. The proposed classifier achieves 95% accuracy using the publicly available standard ASD dataset.

This work presents an investigation on how to define Neural Networks (NN) architectures adopting a data-driven approach using clustering to create sub-labels to facilitate the learning process and to discover the number of neurons needed to compose the layers. We also increase the depth of the model aiming to represent the samples better, the more in-depth it flows into the model. We hypothesize that the clustering process identifies sub-regions in the feature space in which the samples belonging to the same cluster have strong similarities. We used seven benchmark datasets to validate our hypothesis using 10-fold cross validation 3 times. The proposed model increased the performance, while never decreased it, with statistical significance considering the p-value $< 0.05$ in comparison with a Multi-Layer Perceptron with a single hidden layer with approximately the same number of parameters of the architectures found by our approach.

Fundamental frequency has been used for a long time in speaker identification (Braun, 1995; Rose, 2003). The within-speaker variation in F0 is affected by several factors. In Braun (1995), they are categorized as technical, physiological and psychological factors. Tape speed, which surprisingly still is an issue for forensic samples, and sample size are examples of technical factors. Smoking and age are examples of physiological factors, while emotional state and background noise are examples of psychological factors. However, fundamental frequency has been shown to be a successful forensic phonetic parameter (Nolan, 1983). To be able to study differences it is suggested to use long-term distribution measures such as arithmetical mean and standard deviation (Rose, 2002). The duration of the samples should be more than 60 seconds according to Nolan (1983), but Rose (1991) reports that F0 measurements for seven Chinese speakers stabilized much earlier, implying that the values may be ...

In automatic speech recognition (ASR) the non-linear data projection provided by a one hidden layer multilayer perceptron (MLP), trained to recognise phonemes, has previously been shown to provide feature enhancement which can substantially increase ASR performance, especially in noise. Previous attempts to apply an analogous approach to speaker identification have not succeeded in improving performance, except by combining MLP processed features with other features. We present test results for the TIMIT database which show that the advantage of MLP preprocessing for open set speaker identification increases with the number of speakers used to train the MLP and that improved identification is obtained as this number increases beyond sixty. We also present a method for selecting the speakers used for MLP training which further improves identification performance.

We investigate the problem of predicting the quality of automatic speech recognition (ASR) output under the following rigid constraints: i) reference transcriptions are not available, ii) confidence information about the system that produced the transcriptions is not accessible, and iii) training and test data come from multiple domains. To cope with these constraints (typical of the constantly increasing amount of automatic transcriptions that can be found on the Web), we propose a domain-adaptive approach based on multitask learning. Different algorithms and strategies are evaluated with English data coming from four domains, showing that the proposed approach can cope with the limitations of previously proposed single task learning methods.

In this paper, we propose an integration of random subspace sampling and Fishervoice for speaker verification. In the previous random sampling framework [1], we randomly sample the JFA feature space into a set of low-dimensional subspaces. For every random subspace, we use Fishervoice to model the intrinsic vocal characteristics in a discriminant subspace. The complex speaker characteristics are modeled through multiple subspaces. Through a fusion rule, we form a more powerful and stable classifier that can preserve most of the discriminative information. But in many cases, random subspace sampling may discard too much useful discriminative information for high-dimensional feature space. Instead of increasing the number of random subspace or using more complex fusion rules which increase system complexity, we attempt to increase the performance of each individual weak classifier. Hence, we propose to investigate the integration of random subspace sampling with the Fishervoice approa...

Human body characteristics such as fingerprints, retinas and irises, facial struc-ture, and voice recognition are just some of the many biometric fields being researched today. These characteristics are unique to each individual, then they are a good choice for developing biometric keys. ...

The performance of speaker recognition system is highly dependent on the amount of speech used in enrollment and test. This work presents a detailed experimental review and analysis of the GMM-SVM based speaker recognition system in presence of duration variability. This article also reports a comparison of the performance of GMM-SVM classifier with its precursor technique Gaussian mixture model-universal background model (GMM-UBM) classifier in presence of duration variability. The goal of this research work is not to propose a new algorithm for improving speaker recognition performance in presence of duration variability. However, the 2 main focus of this work is on utterance partitioning (UP), a commonly used strategy to compensate the duration variability issue. We have analysed in detailed the impact of training utterance partitioning in speaker recognition performance under GMM-SVM framework. We further investigate the reason why the utterance partitioning is important for boosting speaker recognition performance. We have also shown in which case the utterance partitioning could be useful and where not. Our study has revealed that utterance partitioning does not reduce the data imbalance problem of the GMM-SVM classifier as claimed in earlier study. Apart from these, we also discuss issues related to the impact of parameters such as number of Gaussians, supervector length, amount of splitting required for obtaining better performance in short and long duration test conditions from speech duration perspective. We have performed the experiments with telephone speech from POLYCOST corpus consisting of 130 speakers.

This paper presents statistical data for the fundamental frequency of 100 young male speakers of Standard Southern British English producing spontaneous speech under cognitive stress. The material comes from the new DyViS database, for which subjects underwent a simulated police interview. The distribution of F0 in a large homogeneous group of speakers is of forensic significance since it provides a framework for understanding the significance of F0 measurements in casework. Long-term F0 for the 100 speakers yielded a mode of 102 Hz, a mean of 106 Hz and a median of 105 Hz, and had a near-normal distribution. We demonstrate the limitations of F0 as a discriminatory feature for the majority (60%) of our speaker group, whose long-term F0 occurred within a narrow window of 20 Hz. Conversely, we see the forensic implications for recordings where a speaker’s F0 is outside this window.

Congenital amusia is a neurogenetic disorder affecting musical pitch processing. It also affects lexical tone perception. It is well documented that noisy conditions impact speech perception in second language learners and cochlear implant users. However, it is yet unclear whether and how noise affects lexical tone perception in the amusics. This paper examined the effect of multi-talker babble noise [1] on lexical tone identification and discrimination in 14 Cantonesespeaking amusics and 14 controls at three levels of signal-tonoise ratio (SNR). Results reveal that the amusics were less accurate in the identification of tones compared to controls in all SNR conditions. They also showed degraded performance in the discrimination, but less severe than in the identification. These results confirmed that amusia influences lexical tone processing. But the amusics were not influenced more by noise than the controls in either identification or discrimination. This indicates that the deficits of amusia may not be due to the lack of native-like language processing mechanisms or are mechanical in nature, as in the case of second language learners and cochlear implant users. Instead, the amusics may be impaired in the linguistic processing of native tones, showing impaired tone perception already under the clear condition.

The performance of speaker recognition system is highly dependent on the amount of speech used in enrollment and test. This work presents a detailed experimental review and analysis of the GMM-SVM based speaker recognition system in presence of duration variability. This article also reports a comparison of the performance of GMM-SVM classifier with its precursor technique Gaussian mixture model-universal background model (GMM-UBM) classifier in presence of duration variability. The goal of this research work is not to propose a new algorithm for improving speaker recognition performance in presence of duration variability. However, the 2 main focus of this work is on utterance partitioning (UP), a commonly used strategy to compensate the duration variability issue. We have analysed in detailed the impact of training utterance partitioning in speaker recognition performance under GMM-SVM framework. We further investigate the reason why the utterance partitioning is important for boosting speaker recognition performance. We have also shown in which case the utterance partitioning could be useful and where not. Our study has revealed that utterance partitioning does not reduce the data imbalance problem of the GMM-SVM classifier as claimed in earlier study. Apart from these, we also discuss issues related to the impact of parameters such as number of Gaussians, supervector length, amount of splitting required for obtaining better performance in short and long duration test conditions from speech duration perspective. We have performed the experiments with telephone speech from POLYCOST corpus consisting of 130 speakers.

Important problems in speech soft biometrics include the prediction of speaker's age or gender. Here, the aforementioned problems are addressed in the context of utterances collected during a long time period. A unified framework for age and gender prediction is proposed based on Parallel Factor Analysis 2 (PARAFAC2). PARAFAC2 is applied to a collection of three matrices, namely the speech utterance-feature matrix whose columns are the auditory cortical representations, the speaker age matrix whose columns are indicator vectors of suitable dimension, and the speaker gender matrix whose columns are proper indicator vectors associated to speaker's gender. PARAFAC2 is able to reduce the dimensionality of the auditory cortical representations by projecting these representations onto a semantic space dominated by the age and the gender concepts, yielding a sketch (i.e., a feature vector of reduced dimensions). To predict speaker's age interval associated to a test utterance, the speech utterance sketch is pre-multiplied by the left singular vectors of the speaker age matrix. To predict the gender of the speaker who uttered any test utterance, the speech utterance sketch is pre-multiplied by the left singular vectors of the speaker gender matrix. In both cases, a ranking vector is obtained that is exploited for decision making. Promising results are demonstrated, when the aforementioned framework is applied to the Trinity College Dublin Speaker Ageing Database.

One of the world's chronic neuro-degenerative diseases, Alzheimer's Disease (AD), leads its sufferers, among other symptoms, to suffer from speech difficulties. In particular, the inability to recall vocabulary which makes patients' speech different. Furthermore, Mild Cognitive Impairment (MCI) is usually considered as a prodromal neurodegenerative state of AD. The key to abate the progress of both disorders is their early diagnosis. However, actual ways of diagnosis are costly and quite time-consuming. In this study, we propose the extraction of features from speech through the use of the i-vector approach, by which we seek to model the speech pattern of the three mental conditions from the subjects. To the best of our knowledge, no previous studies have utilized i-vector features to assess Alzheimer's before. These i-vectors are extracted from Mel-Frequency Cepstral Coefficients (MFCCs), then they are given to a SVM classifier in order to identify the speech in one of the following manners: AD-Alzheimer Disease, MCI-Mild Cognitive Impairment, HC-Healthy Control. We tested these i-vector features by performing a 5-fold cross-validation and we achieved an F1-score of 79.2%.

The aim of automatic pathological voice detection systems is to serve as tools, to medical specialists, for a more objective, less invasive and improved diagnosis of diseases. In this respect, the gold standard for those systern^ include the usage of a^optimized representation of the spectral envelope, either based on cepstral coefficients from the mel-scaled Fourier spectral envelope (Mel-Frequency Cepstral Coefficients) or from an all-pole estimation (Linear Prediction Coding Cepstral Coefficients) for characterization, and Gaussian Mixture Models for posterior classification. However, the study of recently proposed GMM-based classifiers as well as Nuisance mitigation techniques, such as those employed in speaker recognition, has not been widely considered in pathology detection labours. The present work aims at testing whether or not the employmenl|of such speaker recognition tools might contribute to improve system performance in pathology detection systems, specifically in the automatic detection of Obstructive Sleep Apnea. The testing procedure employs^arj^Ob3tructivo Sloop Apnoa database^ in conjunction with GMM-based classifiers looking for a better performance. The results show that an improved performance might be obtained by using such approach.

Automatic detection of a baby cry in audio signals is an essential step in applications such as remote baby monitoring. It is also important for researchers, who study the relation between baby cry patterns and various health or developmental parameters. In this paper, we propose two machine-learning algorithms for automatic detection of baby cry in audio recordings. The first algorithm is a low-complexity logistic regression classifier, used as a reference. To train this classifier, we extract features such as Mel-frequency cepstrum coefficients, pitch and formants from the recordings. The second algorithm uses a dedicated convolutional neural network (CNN), operating on log Mel-filter bank representation of the recordings. Performance evaluation of the algorithms is carried out using an annotated database containing recordings of babies (0-6 months old) in domestic environments. In addition to baby cry, these recordings contain various types of domestic sounds, such as parents talking and door opening. The CNN classifier is shown to yield considerably better results compared to the logistic regression classifier, demonstrating the power of deep learning when applied to audio processing.

ASR performance, for current systems, degrades dramatically when there is a mismatch between the training and testing conditions, for instance due to the presence of other sound sources (Lippmann 1997). However, many potential applications will remain unacceptable to users until ASR ...