Automatic Speaker Verification

This paper proposes text independent automatic speaker verification system using IMFCC (Inverse/ Reverse Mel Frequency Coefficients) and ITEM (Information Theoretic Expectation Maximization). To perform speaker verification, feature extraction using Mel scale has been widely applied and has established better results. The IMFCC is based on inverse Mel-scale. The IMFCC effectively captures information available at the high frequency formants which is ignored by the MFCC. In this paper the fusion of MFCC and IMFCC at input level is proposed. GMMs (Gaussian Mixture Models) based on EM (Expectation Maximization) have been widely used for classification of text independent verification. However EM comes across the convergence issue. In this paper we use our proposed ITEM which has faster convergence, to train speaker models. ITEM uses information theory principles such as PDE (Parzen Density Estimation) and KL (Kullback-Leibler) divergence measure. ITEM acclimatizes the weights, means and covariances, like EM. However, ITEM process is not performed on feature vector sets but on a set of centroids obtained using IT (Information Theoretic) metric. The ITEM process at once diminishes divergence measure between PDE estimates of features distribution within a given class and the centroids distribution within the same class. The feature level fusion and ITEM is tested for the task of speaker verification using NIST2001 and NIST2004. The experimental evaluation validates that MFCC/IMFCC has better results than the conventional delta/MFCC feature set. The MFCC/IMFCC feature vector size is also much smaller than the delta MFCC thus reducing the computational burden as well. ITEM method also showed faster convergence, than the conventional EM method, and thus it leads to higher speaker recognition scores.

Speaker verification (SV) is the process of verifying whether speech from two audio signals originate from the same speaker or different speakers. Current state-of-the-art SV systems are based on deep neural networds, predominantly trained using the VoxCeleb dataset. This may lead to varying SV performance when using the models for inference on real-world data. To research these possible variations in performance, three establised SV models, namely the ECAPA-TDNN, ResNet and WavLM, are evaluated on the UCLA variability, CommonVoice, FRIDA and Wyred datasets. The ECAPA-TDNN and ResNet models are found to perform slightly worse when compared with the VoxCeleb evaluation results while the WavLM model performs significantly worse. The ResNet model shows the best performance on all four datasets. After evaluation, the ResNet model is improved by fine-tuning the model on the UCLA dataset and, further by creating a Deep Weight Space Ensemble (WSE) model between the pre-trained and fine-tuned models. Between the pretrained, fine-tuned and WSE models, the WSE model has the best overall performance, attaining the best scores on the UCLA test set. Scores for the other three datasets show a lower decrease than the fine-tuned model. This indicates that fine-tuning with WSE can alleviate the loss in model performance on real-world data.

Speaker modeling distance normalization (D-Norm) is one of the important score normalization techniques in speaker verification (SV) system. For D-Norm implementation, it doesn’t need any additional speech data or external speaker population, that being one of the essential advantages of D-Norm. But traditional D-Norm has still disadvantage of its time complexity due to the Monte-Carlo based Kullback –Leibler distance estimation approach. In this paper we used a simplified an innovative D-Norm technique, which is based on the upper bound of the KL divergence between two statistical models namely GMM-UBM adapted target speaker model and the UBM model. This approach proved better performance. From the experimental point of view we can conclude that the performance of the multilingual SV system has been degrades approximately 3.0% due to language mismatching of the training and testing of the system. Also by applying D-Norm, the performance of the baseline system has been improved appr...

In this paper we report the experiment carried out on recently collected speaker recognition database namely Arunachali Language Speech Database (ALS-DB)to make a comparative study on the performance of acoustic and prosodic features for speaker verification task.The speech database consists of speech data recorded from 200 speakers with Arunachali languages of NorthEast India as mother tongue. The collected database is evaluated using Gaussian mixture model-Universal Background Model (GMM-UBM) based speaker verification system. The acoustic feature considered in the present study is Mel-Frequency Cepstral Coefficients (MFCC) along with its derivatives.The performance of the system has been evaluated for both acoustic feature and prosodic feature individually as well as in combination.It has been observed that acoustic feature, when considered individually, provide better performance compared to prosodic features. However, if prosodic features are combined with acoustic feature, performance of the system outperforms both the systems where the features are considered individually. There is a nearly 5% improvement in recognition accuracy with respect to the system where acoustic features are considered individually and nearly 20% improvement with respect to the system where only prosodic features are considered.

Bibliography: leaves 95-104.Speech recognition systems all have one criterion in common; they perform better in a controlled environment using clean speech. Though performance can be excellent, even exceeding human capabilities for clean speech, systems fail when presented with speech data from more realistic environments such as telephone channels. The differences using a recognizer in clean and noisy environments are extreme, and this causes one of the major obstacles in producing commercial recognition systems to be used in normal environments. It is the lack of performance of speaker recognition systems with telephone channels that this work addresses. The human auditory system is a speech recognizer with excellent performance, especially in noisy environments. Since humans perform well at ignoring noise more than any machine, auditory-based methods are the promising approaches since they attempt to model the working of the human auditory system. These methods have been shown to...

In this paper, we present a novel technique for the estimation of the high frequency components (4-8kHz) of speech signals from narrow-band (0-4 kHz) signals using convolutive Non-Negative Matrix Factorisation (NMF). The proposed technique utilizes a brief recording of simultaneous broad band and narrow band signals from a target speaker to learn a set of broad-band non-negative bases for the speaker. The low-frequency components of these bases are used to determine how the high-frequency components must be combined in order to reconstruct the high-frequency components of new narrow-band signals from the speaker. Experiments reveal that the technique is able to reconstruct broadband sppech that is perceptually virtually indistinguishable from true broadband recordings.

This paper is concerned with the task of speaker verification on audio with multiple overlapping speakers. Most speaker verification systems are designed with the assumption of a single speaker being present in a given audio segment. However, in a real-world setting this assumption does not always hold. In this paper, we demonstrate that current speaker verification systems are not robust against audio with noticeable speaker overlap. To alleviate this issue, we propose marginmixup, a simple training strategy that can easily be adopted by existing speaker verification pipelines to make the resulting speaker embeddings robust against multi-speaker audio. In contrast to other methods, margin-mixup requires no alterations to regular speaker verification architectures, while attaining better results. On our multi-speaker test set based on VoxCeleb1, the proposed margin-mixup strategy improves the EER on average with 44.4% relative to our state-of-the-art speaker verification baseline systems.

Keeping track of the multiple passwords, PINs, memorable dates and other authentication details needed to gain remote access to accounts is one of modern life's less appealing challenges. The employment of a voice-based verification as a biometric technology for both children and adults could be a good replacement to the old fashioned memory dependent procedure. Using voice for authentication could be beneficial in several application areas, including, security, protection, education, call-based and web-based services. Voice-based biometric applications are subject to different types of spoofing attacks. The most accessible and affordable type of spoofing for a voice-based biometrics system is a replay attack. Replay, which is to playback a pre-recorded speech sample, presents a genuine risk to automatic speaker verification technology. This work presents two architectures for detecting frauds caused by replay attacks in a voice-based biometrics authentication systems. Experimental results confirmed that obtained performances from both methods could further improve by applying a machine learning algorithm for performing fusion at the score level. The performance of both methods further improved by fusion using independent sources of scores in different architectures.

Automatic recognition of emotions in speech has attracted the attention of the research community in recent years. Some of the most relevant proposed applications of it are in call-centers. In these scenarios the speech is distorted by compression algorithms. The effects of such distortion on the performance of systems for automatic recognition of emotions must be assessed. In this study these effects are evaluated independently of any other distortions generated by the communications channel. Several stateof-the-art codecs are used to compress the speech signals of two emotional speech databases. The databases used are the Berlin Database of Emotional Speech and the en-terface05. The methodology considers voiced and unvoiced segments of the speech separately. Spectral, cepstral, noise and Non-Linear Dynamics (NLD) measures are used to characterize the segments. Finally, a classifier based on a Gaussian Mixture Model (GMM) is used to identify the emotion. The results indicate that voiced segments are less affected by the compression than unvoiced ones in terms in classification accuracy. They also show that the bandwidth of the analyzed signals is an important factor in the classification results.

This paper introduces an evolutionary approach for training the adaptive network-based fuzzy inference system (ANFIS). The previous works are based on gradient descendent (GD); this algorithm converges very slowly and gets stuck down at bad local minima. This study applies one of the swarm intelligent branches, named particle swarm optimization (PSO), where the premise parameters of the rules are optimized by a PSO, and the conclusion part is optimized by least-squares estimation (LSE). The hybrid PSO-ANFIS model is performed for speaker recognition on CHAINS speech dataset. The results obtained by the hybrid model showed an improvement on the accuracy compared to similar ANFIS based on gradient descendent optimization.

This paper presents new method for improving the security of speaker recognition in case of spoofing attack. In the proposed technique, digital speech watermarking has been applied on speech signal to increase robustness. To achieve this purpose, watermark is embedded in claimed signal at transmission side and then it is sent through the unsecure channel. In receiver side, watermark is extracted as proof of authentication. The results shows that digital speech watermarking can successfully apply for anti-spoofing attack purposes because the quality of the speech signal is not significantly degraded by watermark.

Research on speaker recognition is extending to address the vulnerability in the wild conditions, among which genre mismatch is perhaps the most challenging, for instance, enrollment with reading speech while testing with conversational or singing audio. This mismatch leads to complex and composite inter-session variations, both intrinsic (i.e., speaking style, physiological status) and extrinsic (i.e., recording device, background noise). Unfortunately, the few existing multi-genre corpora are not only limited in size but are also recorded under controlled conditions, which cannot support conclusive research on the multi-genre problem. In this work, we firstly publish CN-Celeb, a large-scale multi-genre corpus that includes in-the-wild speech utterances of 3,000 speakers in 11 different genres. Secondly, using this dataset, we conduct a comprehensive study on the multi-genre phenomenon, in particular the impact of the multi-genre challenge on speaker recognition, and on how to utilize the valuable multi-genre data more efficiently.

Today, the application of artificial neural network tools to define models that mimic the dynamic behavior of structural systems is a widespread approach. A fundamental issue in developing these strategies for damage assessment in civil structures is represented by the unbalanced nature of the available databases, which commonly contain plenty of data coming from the structure under healthy operational conditions and very few samples from the system in unhealthy conditions since the structure would have failed by that time. Consequently, the learning task, carried on with standard deep learning approaches, becomes case-dependent and tends to be specialized for a particular structure and for a very limited number of damage scenarios. This work presents a framework for damage classification in structural systems intended to overcome such limitations. In this methodology, the model is trained to gain knowledge in the learning task from a rich acoustic dataset (source domain), acquiring higher-level features characterizing vibration traits from a rich acoustic dataset. This knowledge is then transferred to a target domain, with much less training data, such as a structural system, in order to classify its structural condition. The framework starts with constructing a Time-Delay Neural Network (TDNN) structure, trained on the VoxCeleb dataset, in the speech domain. The input of the network consists of Cepstral and pitch features extracted from the audio records. Higher-level features, the x-vectors, speaker embeddings, capturing neural outputs of the network's intermediate layers, are derived and then used to train a Probabilistic Linear Discriminant analysis (PLDA) model to provide a probabilistic discriminant model for speaker comparison. These features collect generic information regarding the source domain and characterize a classification process based on the frequency content of signals, which is not strictly dependent on the original acoustic domain. Because of the non-case-dependent nature of the x-vector embeddings (features), they can be used to train an alternative PLDA model to address a damage classification task, considering vibration measurements coming from a different system, a structural one which represents the target domain. The simulated data from the 12 degrees of freedom benchmark shear-building structure provided by the IASC-ASCE Structural Health Monitoring Group are studied to verify the proposed framework's effectiveness.

This paper describes three studies conducted with a total of 114 individuals with hearing loss and 12 hearing controls, with the goal of investigating the impact of audio quality parameters on the accessibility of voice telecommunications. Three categories of parameters are covered: (1) narrowband (NB) versus wideband (WB) audio; (2) encoding audio at varying bit rates, ranging from typical rates used in today's telecom networks to the highest quality supported by these audio codecs; and (3) absence of packet loss to worst-case packet loss in VoIP telephony. With WB audio, individuals with hearing loss exhibit better speech recognition, expend less perceived mental effort, and rate speech quality higher than with NB audio. Bit rate affects speech recognition for NB audio, and speech quality ratings for both NB and WB audio. Packet loss affects all of speech recognition, mental effort, and speech quality ratings. WB versus NB audio also affects hearing individuals, especially under packet loss.

Over the recent years, various self-supervised contrastive embedding learning methods for deep speaker verification were proposed. The performance of the self-supervised contrastive learning framework highly depends on the data augmentation technique, but due to the sensitive nature of speaker information within the speech signal, most speaker embedding training relies on simple augmentations such as additive noise or simulated reverberation. Thus while the conventional self-supervised speaker embedding systems can yield minimum within-utterance variability, the capability to generalize to out-of-set utterance is limited. In order to alleviate this problem, we propose a novel self-supervised learning framework for speaker verification which combines the angular prototypical loss and the instance mix (i-mix) regularization. The proposed method was evaluated on the VoxCeleb1 dataset and showed noticeable improvement over the standard self-supervised embedding method.

Over the recent years, various self-supervised contrastive embedding learning methods for deep speaker verification were proposed. The performance of the self-supervised contrastive learning framework highly depends on the data augmentation technique, but due to the sensitive nature of speaker information within the speech signal, most speaker embedding training relies on simple augmentations such as additive noise or simulated reverberation. Thus while the conventional self-supervised speaker embedding systems can yield minimum within-utterance variability, the capability to generalize to out-of-set utterance is limited. In order to alleviate this problem, we propose a novel self-supervised learning framework for speaker verification which combines the angular prototypical loss and the instance mix (i-mix) regularization. The proposed method was evaluated on the VoxCeleb1 dataset and showed noticeable improvement over the standard self-supervised embedding method.

Voice conversion for speaker anonymization is an emerging concept for privacy protection. In a deep learning setting, this is achieved by extracting multiple features from speech, altering the speaker identity, and waveform synthesis. However, many existing systems do not modify fundamental frequency (F0) trajectories, which convey prosody information and can reveal speaker identity. Moreover, mismatch between F0 and other features can degrade speech quality and intelligibility. In this paper, we formally introduce a method that synthesizes F0 trajectories from other speech features and evaluate its reconstructional capabilities. Then we test our approach within a speaker anonymization framework, comparing it to a baseline and a state-of-the-art F0 modification that utilizes speaker information. The results show that our method improves both speaker anonymity, measured by the equal error rate, and utility, measured by the word error rate.

Keeping track of the multiple passwords, PINs, memorable dates and other authentication details needed to gain remote access to accounts is one of modern life's less appealing challenges. The employment of a voice-based verification as a biometric technology for both children and adults could be a good replacement to the old fashioned memory dependent procedure. Using voice for authentication could be beneficial in several application areas, including, security, protection, education, call-based and web-based services. Voice-based biometric applications are subject to different types of spoofing attacks. The most accessible and affordable type of spoofing for a voice-based biometrics system is a replay attack. Replay, which is to playback a pre-recorded speech sample, presents a genuine risk to automatic speaker verification technology. This work presents two architectures for detecting frauds caused by replay attacks in a voice-based biometrics authentication systems. Experimental results confirmed that obtained performances from both methods could further improve by applying a machine learning algorithm for performing fusion at the score level. The performance of both methods further improved by fusion using independent sources of scores in different architectures.

Speaker recognition is challenging in real-world environments. Typically, studies approach noises only in an additive manner. However, real environments commonly present reverberating conditions that worsen speech processing. When not considering reverberation in the system modeling, the system may not be robust when applied to real-world conditions. In this work, we use a slight different approach to simulate reverberation, considering randomized conditions of the environment. With this approach, each VoxCeleb1 test sample is corrupted by randomly generated conditions, with diversified amplitudes of noise and speech. We generate a corrupted dataset, in which the best model EER degraded from 0.93% to 30.13%. To improve this degradation, we propose using Normalized Kernel Least-Mean-Square (NKLMS) adaptive filter. Through the use of NKLMS, we were able to improve the EER from 30.13% to 1.11%. The results indicate that NKLMS has a great potential for speech enhancement to improve speaker recognition.

Text categorization is widely used when organizing documents in a digital form. Due to the increasing number of documents in digital form, automated text categorization has been emerged as an appropriate tool to classify documents into predefined categories. High dimensionality of the feature space is a common problem in text categorization. Most of the features affecting the classifier performance are irrelevant and redundant. Hence, feature selection is used to reduce feature space thus increasing classifier performance. In this paper, a twostage method is proposed for text feature selection. At the first stage a filtering technique using the fuzzy entropy measure is applied and features are ranked based on their values. Then, features with the values higher than a threshold are removed from feature set. In the second stage, an ant colony optimization approach is employed to select features from the reduced feature space in the first stage. The proposed method is evaluated through the use of the k-nearest neighbor classifier on top 10 Retures-21578 categories. The experimental results obtained, show the efficiency of the proposed method.

In this paper we present an efficient method for training models for speaker recognition using small or under-resourced datasets. This method requires less data than other SOTA (State-Of-The-Art) methods, e.g. the Angular Prototypical and GE2E loss functions, while achieving similar results to those methods. This is done using the knowledge of the reconstruction of a phoneme in the speaker's voice. For this purpose, a new dataset was built, composed of 40 male speakers, who read sentences in Portuguese, totaling approximately 3h. We compare the three best architectures trained using our method to select the best one, which is the one with a shallow architecture. Then, we compared this model with the SOTA method for the speaker recognition task: the Fast ResNet-34 trained with approximately 2,000 hours, using the loss functions Angular Prototypical and GE2E. Three experiments were carried out with datasets in different languages. Among these three experiments, our model achieved the second best result in two experiments and the best result in one of them. This highlights the importance of our method, which proved to be a great competitor to SOTA speaker recognition models, with 500x less data and a simpler approach.

Although speaker verification has conventionally been an audio-only task, some practical applications provide both audio and visual streams of input. In these cases, the visual stream provides complementary information and can often be leveraged in conjunction with the acoustics of speech to improve verification performance. In this study, we explore audiovisual approaches to speaker verification, starting with standard fusion techniques to learn joint audiovisual (AV) embeddings, and then propose a novel approach to handle cross-modal verification at test time. Specifically, we investigate unimodal and concatenation based AV fusion and report the lowest AV equal error rate (EER) of 0.7% on the VoxCeleb1 dataset using our best system. As these methods lack the ability to do cross-modal verification, we introduce a multi-view model which uses a shared classifier to map audio and video into the same space. This new approach achieves 28% EER on VoxCeleb1 in the challenging testing condition of cross-modal verification.

For a few years, the problem of session variability in textindependent automatic speaker verification is being tackled actively. A new paradigm based on a factor analysis model have successfully been applied for this task. While very efficient, its implementation is demanding. In this paper, the algorithms involved in the eigenchannel MAP model are written down for a straightforward implementation, without referring to previous work or complex mathematics. In addition, a different compensation scheme is proposed where the standard GMM likelihood can be used without any modification to obtain good performance (even without the need of score normalization). The use of the compensated supervectors within a SVM classifier through a distance based kernel is also investigated. Experiments results shows an overall 50% relative gain over the standard GMM-UBM system on NIST SRE 2005 and 2006 protocols (both at the DCFmin and EER).

As acoustic signal generated from vocal folds is directly affected by vocal tract pathologies, it can be an effective tool for diagnosis purpose. In this work, we present an efficient method for voice pathology detection based on speech signal processing and machine learning techniques. In the proposed method, we used MFCC to represent the signal features, and we chose to combine GMM and SVM classifiers to benefit from their generative and discriminative natures respectively. That is to exploit the similarity function of the RBF kernel to separate the GMM models representing normal and pathological voices. To further improve the separation, we used modified versions of the well known Kullback-leibler and Bhattacharyya distances. The modified distances, unlike the classical ones, do satisfy all metric axioms. As a result, we obtained an improvement of 2 % and 4 % in terms of sensitivity compared to using the classical Kullback-leibler and Bhattacharyya distances respectively. The Receiver Operating Curve (ROC) does illustrate the efficiency of the proposed method.

The embedding-based deep convolution neural networks (C-NNs) have demonstrated effective for text-independent speaker verification systems with short utterances. However, the duration robustness of the existing deep CNNs based algorithms has not been investigated when dealing with utterances of arbitrary duration. To improve robustness of embedding-based deep C-NNs for longer duration utterances, we propose a novel algorithm to learn more discriminative utterance-level embeddings based on the Inception-ResNet speaker classifier. Specifically, the discriminability of embeddings is enhanced by reducing intra-speaker variation with center loss, and simultaneously increasing inter-speaker discrepancy with softmax loss. To further improve system performance when long utterances are available, at test stage long utterances are segmented into shorter ones, where utterance-level speaker embeddings are extracted by an average pooling layer. Experimental results show that when cosine distance is employed as the measure of similarity for a trial, the proposed method outperforms ivector/PLDA framework for short utterances and is effective for long utterances.

This study proposes a novel scheme of syllable-dependent discriminative speaker embedding learning for small footprint text-dependent speaker verification systems. To suppress undesired syllable variation and enhance the power of discrimination inherited in the frame-level features, we design a novel syllable-dependent clustering loss to optimize the network. Specifically, this loss function utilizes syllable labels as auxiliary supervision information to explicitly maximize intersyllable divisibility and intra-syllable compactness between the learned frame-level features. Successively, we propose two syllable-dependent pooling mechanisms to aggregate the frame-level features to several syllable-level features by averaging those features corresponding to each syllable. The utterance-level speaker embeddings with powerful discrimination are then obtained by concatenating the syllable-level features. Experimental results on Tencent voice wake-up dataset show that our proposed scheme can accelerate the network convergence and achieve significant performance improvement against the state-of-the-art methods.

The embedding-based deep convolution neural networks (C-NNs) have demonstrated effective for text-independent speaker verification systems with short utterances. However, the duration robustness of the existing deep CNNs based algorithms has not been investigated when dealing with utterances of arbitrary duration. To improve robustness of embedding-based deep C-NNs for longer duration utterances, we propose a novel algorithm to learn more discriminative utterance-level embeddings based on the Inception-ResNet speaker classifier. Specifically, the discriminability of embeddings is enhanced by reducing intra-speaker variation with center loss, and simultaneously increasing inter-speaker discrepancy with softmax loss. To further improve system performance when long utterances are available, at test stage long utterances are segmented into shorter ones, where utterance-level speaker embeddings are extracted by an average pooling layer. Experimental results show that when cosine distance is employed as the measure of similarity for a trial, the proposed method outperforms ivector/PLDA framework for short utterances and is effective for long utterances.

Introduction The literature reports on high-frequency audiometry as one of the exams used on hearing monitoring of individuals exposed to high sound pressure in their work environment, due to the method s greater sensitivity in early identification of hearing loss caused by noise. The frequencies that compose the exam are generally between 9 KHz and 20KHz, depending on the equipment. Objective This study aims to perform a retrospective and secondary systematic revision of publications on high-frequency audiometry on hearing monitoring of individuals exposed to occupational noise. Data Synthesis This systematic revision followed the methodology proposed in the Cochrane Handbook, focusing on the question: "Is High-frequency Audiometry more sensitive than Conventional Audiometry in the screening of early hearing loss individuals exposed to occupational noise?" The search was based on PubMed data, Base, Web of Science (Capes), Biblioteca Virtual em Saúde (BVS), and in the references cited in identified and selected articles. The search resulted in 6059 articles in total. Of these, only six studies were compatible with the criteria proposed in this study. Conclusion The performed meta-analysis does not definitively answer the study's proposed question. It indicates that the 16 KHz high frequency audiometry (HFA) frequency is sensitive in early identification of hearing loss in the control group (medium difference (MD ¼ 8.33)), as well as the 4 KHz frequency (CA), this one being a little less expressive (MD ¼ 5.72). Thus, others studies are necessary to confirm the HFA importance for the early screening of hearing loss on individuals exposed to noise at the workplace.

In this study, the researcher investigates one of the testing procedures on acquiring the unknown sampling frequency which is relevant in (ADC) Analog-to-Digital Conversion on Digital Signal Processing. With the help of the Simulink tool in MATLAB, the researcher could be able to conclude after numerous simulations run attempts. The simulation result shows that to avoid aliasing phenomenon, the sampling frequency must be at least twice its maximum frequency.

The impact of speech compression in the automatic classification of speakers with Parkinson's disease (PD) and healthy controls (HC) is tested. The set of codecs considered to compress the speech recordings includes G.722, G.226, GSM-EFR, AMR-WB, SILK, and Opus. A total of 100 speakers (50 with PD and 50 HC) are asked to read a text with 36 words. The recordings are compressed from bit-rates of 705.6 kbps down to 6.6 kbps. The method addressed to discriminate between speakers with PD and HC consists on the systematic segmentation of voiced and unvoiced speech frames. Each kind of frame is characterized independently. For voiced segments noise, perturbation, and cepstral features are considered. The unvoiced segments are characterized with Bark band energies and cepstral features. According to the results the codecs evaluated in this paper do not affect significantly the accuracy of the system, indicating that the addressed methodology could be used for the telemonitoring of PD patients through Internet or through the mobile communications network.

The purpose of this study is to discuss the design and implementation of autonomous surface vehicle (ASV) systems. There’s a lot riding on the advancement and improvement of ASV applications, especially given the benefits they provide over other biometric approaches. Modern speaker recognition systems rely on statistical models like hidden Markov model (HMM), support vector machine (SVM), artificial neural networks (ANN), generalized method of moments (GMM), and combined models to identify speakers. Using a French dataset, this study investigates the effectiveness of prompted te xt speaker verification. At a context-free, single mixed mono phony level, this study has been constructing a continuous speech system based on HMM. After that, suitable voice data is used to build the client and world models. In order to verify speakers, the text-dependent speaker ver-ification system uses sentence HMM that have been concatenated for the key text. Normalized log-likelihood is determined fro...

We aim to describe different approaches for vector quantization in Automatic Speaker Verification. We designed our novel architecture based on multiples codebook representing the speakers and the impostor model called universal background model and compared it to another vector quantization approach used for reducing training data. We compared our scheme with the baseline system, Gaussian Mixtures Models and Maximum a Posteriori Adaptation. The present study demonstrates that the multiples codebook gives more verification accuracy called equal error rate but this improvement also depends on the codebook size.

State-of-the-art text-independent speaker recognition systems for long recordings (a few minutes) are based on deep neural network (DNN) speaker embeddings. Current implementations of this paradigm use short speech segments (a few seconds) to train the DNN. This introduces a mismatch between training and inference when extracting embeddings for long duration recordings. To address this, we present a DNN refinement approach that updates a subset of the DNN parameters with full recordings to reduce this mismatch. At the same time, we also modify the DNN architecture to produce embeddings optimized for cosine distance scoring. This is accomplished using a largemargin strategy with angular softmax. Experimental validation shows that our approach is capable of producing embeddings that achieve record performance on the SITW benchmark.

This paper describes the systems developed by the JHU team for the speaker recognition track of the 2019 VOiCES from a Distance Challenge. On this far-field task, we achieved good performance using systems based on state-of-the-art deep neural network (DNN) embeddings. In this paradigm, a DNN maps variable-length speech segments to speaker embeddings, called x-vectors, that are then classified using probabilistic linear discriminant analysis (PLDA). Our submissions were composed of three x-vector-based systems that differed primarily in the DNN architecture, temporal pooling mechanism, and training objective function. On the evaluation set, our best single-system submission used an extended time-delay architecture, and achieved 0.435 in actual DCF, the primary evaluation metric. A fusion of all three x-vector systems was our primary submission, and it obtained an actual DCF of 0.362.

Large-scale mobile communication systems tend to contain legacy transmission channels with narrowband bottlenecks, resulting in characteristic 'telephone-quality' audio. While higher quality codecs exist, due to the scale and heterogeneity of the networks, transmitting higher sample rate audio with modern high-quality audio codecs can be difficult in practice. This paper proposes an approach where a communication node can instead extend the bandwidth of a band-limited incoming speech signal that may have been passed through a low-rate codec. To this end, we propose a WaveNet-based model conditioned on a log-mel spectrogram representation of a bandwidth-constrained speech audio signal of 8 kHz and audio with artifacts from GSM full-rate (FR) compression to reconstruct the higher-resolution signal. In our experimental MUSHRA evaluation, we show that a model trained to upsample to 24kHz speech signals from audio passed through the 8kHz GSM-FR codec is able to reconstruct audio only slightly lower in quality to that of the Adaptive Multi-Rate Wideband audio codec (AMR-WB) codec at 16kHz, and closes around half the gap in perceptual quality between the original encoded signal and the original speech sampled at 24kHz. We further show that when the same model is passed 8kHz audio that has not been compressed, is able to again reconstruct audio of slightly better quality than 16kHz AMR-WB, in the same MUSHRA evaluation.

The effectiveness of introducing deep neural networks into conventional speaker recognition pipelines has been broadly shown to benefit system performance. A noveltext-independent speaker verification (SV) framework based on the triplet loss and a very deep convolutional neuralnetwork architecture (i.e., Inception-Resnet-v1) are investigated in this study, where a fixedlength speaker discriminative embedding is learned from sparse speech features and utilized as a feature representation for the SV tasks. A concise description of the neural network based speaker discriminative training with triplet loss is presented. An Euclidean distance similarity metric is applied in both network training and SV testing, which ensures the SV system to follow an end-to-end fashion. By replacing the final max/average pooling layer with a spatial pyramid pooling layer in the Inception-Resnet-v1 architecture, the fixed-length input constraint is relaxed and an obvious performance gain is achieved compared with the fixed-length input speaker embedding system. For datasets with more severe training/test condition mismatches, the probabilistic linear discriminant analysis (PLDA) back end is further introduced to replace the distance based scoring for the proposed speaker embedding system. Thus, we reconstruct the SV task with a neural network based front-end speaker embedding system and a PLDA that provides channel and noise variabilities compensation in the back end. Extensive experiments are conducted to provide useful hints that lead to a better testing performance. Comparison with the state-of-theart SV frameworks on three public datasets (i.e., a prompt speech corpus, a conversational speech Switchboard corpus, and NIST SRE10 10 s-10 s condition) justifies the effectiveness of our proposed speaker embedding system. Index Terms-Speaker recognition, very deep convolutional neutral networks, i-vector, PLDA, triplet loss, spatial pyramid pooling. I. INTRODUCTION S PEAKER verification (SV) is a binary classification problem which aims to verify a claimed identify based on the claimed/enrolled speaker model. According to different Manuscript

Forensic investigation often uses biometric evidence as important aids for identifying the culprits. Speech is one of the easily available biometrics in today's hi-tech world. But, most of the speech biometric evidence acquired for investigative purposes will usually be highly distorted. Among these distortions, most prominent is the distortion introduced by the speech codec. Speech codec may either remove or distort some of the speaker-specific features, and this may reduce the speaker verification accuracy. The effect of distortion on commonly used speaker-specific features namely Mel Frequency Cepstral Coefficients (MFCC) and Power Normalized Cepstral Coefficients (PNCC), due to Code Excited Linear Prediction (CELP) codec (the most widely used speech codec in today's mobile telephony), is quantified in this paper. The features which are least affected by the codec are experimentally determined as PNCC. But, when these PNCC coefficients are directly employed, speaker verif...

Feature selection is an important preprocessing step in machine learning and pattern recognition. The ultimate goal of feature selection is to select a feature subset from the original feature set to increase the performance of learning algorithms. In this paper a novel feature selection method based on the graph clustering approach and ant colony optimization is proposed for classification problems. The proposed method's algorithm works in three steps. In the first step, the entire feature set is represented as a graph. In the second step, the features are divided into several clusters using a community detection algorithm and finally in the third step, a novel search strategy based on the ant colony optimization is developed to select the final subset of features. Moreover the selected subset of each ant is evaluated using a supervised filter based method called novel separability index. Thus the proposed method does not need any learning model and can be classified as a filter based feature selection method. The proposed method integrates the community detection algorithm with a modified ant colony based search process for the feature selection problem. Furthermore, the sizes of the constructed subsets of each ant and also size of the final feature subset are determined automatically. The performance of the proposed method has been compared to those of the state-of-the-art filter and wrapper based feature selection methods on ten benchmark classification problems. The results show that our method has produced consistently better classification accuracies.

In this paper we propose a digit-based text-dependent speaker verification system (SVS) in Spanish. The system uses word level Hidden Markov Models (HMM) as classifiers and Frequency Cepstral Coefficients (MFCC) with Cepstral Mean Subtraction as features. The system was developed considering a gender independent Universal Background Model (UBM) within an HMM-UBM framework. The training data set was pooled with both genders equally represented for the UBM. A Target Speaker Model (TSM) was generated using Maximum A Posteriori (MAP) to adapt the UBM’s parameters to the acoustic characteristics of the target speakers. Every target speaker (TS) has a 4 digits password. Robustness of the system was tested using adaptive noise cancellation as a speech enhancement scheme; the speech signals were corrupted with additive white Gaussian noise (AWGN) at different values of signal to noise ratio. Generation of a digit-database, which we have named as BIOMEX-DB, is described. The main contributio...

Speaker verification is an essential task in speech processing with great authentication and surveillance applications. Large-scale datasets have hugely contributed to the success of neural networks for speaker verification. However, in low-resource languages, building such massive datasets is infeasible. This paper aims at proposing a speaker verification model for low-resource scenarios, with the baseline from Clova’s H/ASP system in VoxSRC 2020. The proposed method adopts transfer learning to utilize the knowledge from the model trained with VoxCeleb, an English large-scale dataset. For network optimization, Stochastic Gradient Descent is employed instead of Adam because of its superior generalization. This work also proposes a novel marginal variant of Angular Prototypical (AP) Loss, i.e. Angular Margin Prototypical (AMP) Loss, which encourages a more discriminative embedding space. To experiment with the proposed model, we investigated building a public speaker verification dat...

Automatic speaker verification (ASV) is one of the most natural and convenient means of biometric person recognition. Unfortunately, just like all other biometric systems, ASV is vulnerable to spoofing, also referred to as "presentation attacks." These vulnerabilities are generally unacceptable and call for spoofing countermeasures or "presentation attack detection" systems. In addition to impersonation, ASV systems are vulnerable to replay, speech synthesis, and voice conversion attacks. The ASVspoof challenge initiative was created to foster research on anti-spoofing and to provide common platforms for the assessment and comparison of spoofing countermeasures. The first edition, ASVspoof 2015, focused upon the study of countermeasures for detecting of text-to-speech synthesis (TTS) and voice conversion (VC) attacks. The second edition, ASVspoof 2017, focused instead upon replay spoofing attacks and countermeasures. The ASVspoof 2019 edition is the first to consider all three spoofing attack types within a single challenge. While they originate from the same source database and same underlying protocol, they are explored in two specific use case scenarios. Spoofing attacks within a logical access (LA) scenario are generated with the latest speech synthesis and voice conversion technologies, including state-of-the-art neural acoustic and waveform model techniques. Replay spoofing attacks within a physical access (PA) scenario are generated through carefully controlled simulations that support much more revealing analysis than possible previously. Also new to the 2019 edition is the use of the tandem detection cost function metric, which reflects the impact of spoofing and countermeasures on the reliability of a fixed ASV system. This paper describes the database design, protocol, spoofing attack implementations, and baseline ASV and countermeasure results. It also describes a human assessment on spoofed data in logical access. It was demonstrated that the spoofing data in the ASVspoof 2019 database have varied degrees of perceived quality and similarity to the target speakers, including spoofed data that cannot be differentiated from bona fide utterances even by human subjects. It is expected that the ASVspoof 2019 database, with its varied coverage of different types of spoofing data, could further foster research on anti-spoofing.

State-of-the-art speaker recognition systems performance degrades considerably in noisy environments even though they achieve very good results in clean conditions. In order to deal with this strong limitation, we aim in this work to remove the noisy part of an i-vector directly in the i-vector space. Our approach offers the advantage to operate only at the i-vector extraction level, letting the other steps of the system unchanged. A maximum a posteriori (MAP) procedure is applied in order to obtain clean version of the noisy i-vectors taking advantage of prior knowledge about clean i-vectors distribution. To perform this MAP estimation, Gaussian assumptions over clean and noise i-vectors distributions are made. Operating on NIST 2008 data, we show a relative improvement up to 60% compared with baseline system. Our approach also outperforms the "multi-style" backend training technique. The efficiency of the proposed method is obtained at the price of relative high computational cost. We present at the end some ideas to improve this aspect.

Speaker verification has recently been introduced to the forensic field as a new and complimentary approach to other forensic methods. With the advancement in speech communication technologies including voice over IP and wireless multimedia applications, speech is seldom sent between two parties in plain, it is at least partially encrypted before transmission. We present automatic speaker verification techniques based on hidden Markov and Gaussian mixture models from partially encrypted speech from the perceptually less relevant speech features which are unencrypted. An equal error rate (EER) of 23% and minimum detection cost value of 8% has been achieved on a database of 84 speakers using adapted Gaussian mixture modeling. Comparison between different modeling techniques and effect of Gaussian mixture densities are also carried out and results are tabulated. The results suggest that partial or selective encryption techniques may provide content protection but will not protect the speaker's identity.

"C-Tio.br rogeroclaritytechinc.com Absfrm-Noise robustness of automatic speaker verification systems is crucial in real life applications. A study on the performance of several spectral subtraction-based speech enhancement techniques shows the poor performance of these algorithms when used as a preprocessing stage of the speaker verification system in the presence of colored noise. In this paper, a new technique based on the addition of modeled colored noise is introduced. Experimental results in both white and colored noise environments are presented, showing the improvement of the proposed scheme.

Speech is composed of different sounds (acoustic segments). Speakers differ in their pronunciation of these sounds. The segmental approaches described in this paper are meant to exploit these differences for speaker verification purposes. For such approaches, the speech is divided into different classes, and the speaker modeling is done for each class. The speech segmentation applied is based on automatic language independent speech processing tools that provide a segmentation of the speech requiring neither phonetic nor orthographic transcriptions of the speech data. Two different speaker modeling approaches, based on multilayer perceptrons (MLPs) and on Gaussian mixture models (GMMs), are studied. The MLPbased segmental systems have performance comparable to that of the global MLP-based systems, and in the mismatched train-test conditions slightly better results are obtained with the segmental MLP system. The segmental GMM systems gave poorer results than the equivalent global GMM systems.

Acoustic noise is a big challenge for speaker recognition systems. The state-of-the-art speaker recognition systems are based on deep neural network speaker embeddings called xvector extractor. A noise-robust x-vector extractor is highly demanded in speaker recognition systems. In this paper, we introduce Barlow Twins self-supervised loss function in the area of speaker recognition. Barlow Twins objective function tries to optimize two criteria: Firstly, it increases the similarity between two versions of the same signal (i.e. the clean and its augmented noisy version) to make the speaker embedding invariant to the acoustic noise. Secondly, it reduces the redundancy between dimensions of the x-vectors that improves the overall quality of speaker embeddings. In our research, Barlow Twins objective function is integrated with the ResNet-based speaker embedding system. In the proposed system, the Barlow Twins objective function is calculated in the embedding layer and it is optimized jointly with the speaker classifier loss function. The experimental results on Fabiole corpus show 22 % relative gain in terms of EER in the clean environments and 18% improvement in the presence of noise with low SNR and reverberation.