Gaussian Mixture Model-Universal Background Model Research Articles

Development of efficient techniques for ASR System for Speech Detection and Recognization system using Gaussian Mixture Model- Universal Background Model

Some practical uses of ASR have been implemented, including the transcription of meetings and the usage of smart speakers. It is the process by which speech waves are transformed into text that allows computers to interpret and act upon human speech. Scalable strategies for developing ASR systems in languages where no voice transcriptions or pronunciation dictionaries exist are the primary focus of this work. We first show that the necessity for voice transcription into the target language can be greatly reduced through cross-lingual acoustic model transfer when phonemic pronunciation lexicons exist in the new language. Afterwards, we investigate three approaches to dealing with languages that lack a pronunciation lexicon. Secondly, we have a look at the efficiency of graphemic acoustic model transfer, which makes it easy to build pronunciation dictionaries. Thesis problems can be solved, in part, by investigating optimization strategies for training on huge corpora (such as GA+HMM and DE+HMM). In the training phase of acoustic modelling, the suggested method is applied to traditional methods. Read speech and HMI voice experiments indicated that while each data augmentation strategy alone did not always increase recognition performance, using all three techniques together did. Power normalised cepstral coefficient (PNCC) features are tweaked somewhat in this work to enhance verification accuracy. To increase speaker verification accuracy, we suggest employing multiple “Gaussian Mixture Model-Universal Background Model (GMM-UBM) and SVM classifiers”. Importantly, pitch shift data augmentation and multi-task training reduced bias by more than 18% absolute compared to the baseline system for read speech, and applying all three data augmentation techniques during fine tuning reduced bias by more than 7% for HMI speech, while increasing recognition accuracy of both native and non-native Dutch speech.

Influence of tree-based multi-layer node information on scoring accuracy and speed of speaker verification

In the speaker verification task based on Gaussian Mixture Model-Universal Background Model (GMM-UBM), by constructing the UBM as a tree structure, the kernel Gaussians suitable for different speakers can be quickly selected, which speeds up the modeling of speaker acoustic space by GMM. The tree-based kernel selection algorithm (TBKS) introduces a beam-width, which increases the candidate range of kernels and improves the kernel selection accuracy. In this paper, we improve the TBKS algorithm by introducing a recall rate to adjust the number of nodes recalled in each layer of the tree structure. This adjustment refines the quantity and resolution of Gaussian distributions in various subspaces within the acoustic space, compensating for the loss caused by discarding some significant Gaussians erroneously. Speaker verification experiments are carried out based on the Aishell2 dataset. The results reveal that the modified TBKS algorithm reduces EER by 7.5% relatively and increses computational reduction factor to 42.93, enhancing both recognition accuracy and speed. In addition, the test speech is spliced into different lengths and common environmental noise is added to verify the universality of the improved algorithm.

On Training Targets and Activation Functions for Deep Representation Learning in Text-Dependent Speaker Verification

Deep representation learning has gained significant momentum in advancing text-dependent speaker verification (TD-SV) systems. When designing deep neural networks (DNN) for extracting bottleneck (BN) features, the key considerations include training targets, activation functions, and loss functions. In this paper, we systematically study the impact of these choices on the performance of TD-SV. For training targets, we consider speaker identity, time-contrastive learning (TCL), and auto-regressive prediction coding, with the first being supervised and the last two being self-supervised. Furthermore, we study a range of loss functions when speaker identity is used as the training target. With regard to activation functions, we study the widely used sigmoid function, rectified linear unit (ReLU), and Gaussian error linear unit (GELU). We experimentally show that GELU is able to reduce the error rates of TD-SV significantly compared to sigmoid, irrespective of the training target. Among the three training targets, TCL performs the best. Among the various loss functions, cross-entropy, joint-softmax, and focal loss functions outperform the others. Finally, the score-level fusion of different systems is also able to reduce the error rates. To evaluate the representation learning methods, experiments are conducted on the RedDots 2016 challenge database consisting of short utterances for TD-SV systems based on classic Gaussian mixture model-universal background model (GMM-UBM) and i-vector methods.

Power Normalized Gammachirp Cepstral (PNGC) coefficients-based approach for robust speaker recognition

Speaker identification or recognition task aims to identify persons from their voices. This paper introduces a new feature extraction approach for robust speaker recognition named Power Normalized Gammachirp Cepstral (PNGC). Key aspect of our method is the use of a biologically motivated auditory perceptual model. For speaker modeling, we use the Gaussian Mixture Model-Universal Background Model (GMM-UBM). The proposed approach includes the following main processing steps: (1) an efficient auditory filter model based on the normalized Gammachirp auditory Filterbank to estimate the cochlea spectral behavior, (2) an environmental noise compensation bloc that employs: a medium-time power analysis, an asymmetric noise-suppression module with a temporal masking module, and a frequency smoothing module to compensate the environmental noise effects, and (3) a power-law nonlinearity. Conducted experimentations on TIMIT and Aurora datasets proved that our proposed PNGC approach achieves an improved recognition accuracy compared to the MFCC and recent PNCC methods for noisy speaker recognition.

Familiar and unfamiliar speaker recognition assessment and system emulation for cochlear implant users.

In the area of speech processing, human speaker identification under naturalistic environments is a challenging task, especially for hearing-impaired individuals with cochlear implants (CIs) or hearing aids (HAs). Motivated by the fact that electrodograms reflect direct CI stimulation of input audio, this study proposes a speaker identification (ID) investigation using two-dimensional electrodograms constructed from the responses of a CI auditory system to emulate CI speaker ID capabilities. Features are extracted from electrodograms through an identity vector (i-vector) framework to train and generate identity models for each speaker using a Gaussian mixture model-universal background model followed by probabilistic linear discriminant analysis. To validate the proposed system, perceptual speaker ID for 20 normal hearing (NH) and seven CI listeners was evaluated with a total of 41 different speakers and compared with the scores from the proposed system. A one-way analysis of variance showed that the proposed system can reliably predict the speaker ID capability of CI (F[1,10] = 0.18, p = 0.68) and NH (F[1,20] = 0, p = 0.98) listeners in naturalistic environments. The impact of speaker familiarity is also addressed, and the results show a reduced performance for speaker recognition by CI subjects using their CI processor, highlighting limitations of current speech processing strategies used in CIs/HAs.

Joint short-time speaker recognition and tracking using sparsity-based source detection

A random finite set-based sequential Monte–Carlo tracking method is proposed to track multiple acoustic sources in indoor scenarios. The proposed method can improve tracking performance by introducing recognized speaker identities from the received signals. At the front-end, the degenerate unmixing estimation technique (DUET) is employed to separate the mixed signals, and the time delay of arrival (TDOA) is measured. In addition, a criterion to select the reliable microphone pair is designed to quickly obtain accurate speaker identities from the mixed signals, and the Gaussian mixture model universal background model (GMM-UBM) is employed to train the speaker model. In the tracking step, the update of the weight for each particle is derived after introducing the recognized speaker identities, which results in better association between the measurements and sources. Simulation results demonstrate that the proposed method can improve the accuracy of the filter states and discriminate the sources close to each other.

Depression assessment in people with Parkinson’s disease: The combination of acoustic features and natural language processing

Parkinson’s disease produces motor impairments such as bradykinesia, rigidity, and different speech impairments, as same as non-motor symptoms like cognitive decline and depression disturbances. Most studies are focused on the analysis of motor symptoms, and only few works study non-motor impairments. Depression is one of the typical non-motor symptoms developed by many Parkinson’s patients. Impairments in speech production together with depression produce negative effects in the communication capabilities and social interaction of patients. This study proposes a combination of speech analysis and natural language processing methods to extract features from spontaneous speech utterances and their transcripts. We consider state-of-the-art word-embedding methods like Bidirectional Encoder Representations from Transformer (BERT) to process the transcripts, and traditional acoustic features such as Bark band energies and Mel frequency cepstral coefficients to model the speech signals. The features are processed with supervectors generated by Gaussian Mixture Model-Universal Background Model (GMM-UBM) and Support Vector Machine (SVM) classifiers. The dataset consists of 60 Parkinson’s patients divided into two classes according to the depression item of the MDS-UPDRS. The automatic classification of depressed and non-depressed Parkinson’s patients showed F-scores of up to 0.77, which confirms that acoustic and linguistic information embedded in language production can be used for depression analysis in Parkinson’s patients. We present one of the few studies that evaluates depression in Parkinson’s patients considering the combination of acoustic and linguistic information.

Application of Artificial Intelligence on Post Pandemic Situation and Lesson Learn for Future Prospects

ABSTRACT Coronavirus disease (COVID-19) pandemic has intensively damaged human socio-economic lives and the growth of countries around the world. Many efforts have been made in the direction of artificial intelligence (AI) techniques to detect the corona at an early stage and take necessary precautions to stop it from spreading or recovery from the infection. However, the situation and solutions are still challenging. In this paper, we proposed various technological aspects, solutions using a supervised/unsupervised manner and continuous health monitoring with physiological parameters. Finally, the performance of COVID-19 detection with Gaussian mixture model-universal background model (GMM-UBM) technique using the voice signal has been demonstrated. The developed system achieves the COVID-19 detection performance in terms of areas under receiver operating characteristic (ROC) curves in the range 60–67%. Moreover, the various lessons learned from the current COVID-19 crisis are presented for future directions.

Utterance partitioning for speaker recognition: an experimental review and analysis with new findings under GMM-SVM framework

The performance of speaker recognition system is highly dependent on the duration 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 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.

A Comparative Assessment of Text-independent Automatic Speaker Identification Methods Using Limited Data

Automatic Speaker Identification (ASI) is one of the active fields of research in signal processing. Various machine learning algorithms have been used for this purpose. With the recent developments in hardware technologies and data accumulation, Deep Learning (DL) methods have become the new state-of-the-art approach in several classification and identification tasks. In this paper, we evaluate the performance of traditional methods such as Gaussian Mixture Model-Universal Background Model (GMM-UBM) and DL-based techniques such as Factorized Time-Delay Neural Network (FTDNN) and Convolutional Neural Networks (CNN) for text-independent closed-set automatic speaker identification on two datasets with different conditions. LibriSpeech is one of the experimental datasets, which consists of clean audio signals from audiobooks, collected from a large number of speakers. The other dataset was collected and prepared by us, which has rather limited speech data with low signal-to-noise-ratio from real-life conversations of customers with the agents in a call center. The duration of the speech signals in the query phase is an important factor affecting the performances of ASI methods. In this work, a CNN architecture is proposed for automatic speaker identification from short speech segments. The architecture design aims at capturing the temporal nature of speech signal in an optimum convolutional neural network with low number of parameters compared to the well-known CNN architectures. We show that the proposed CNN-based algorithm performs better on the large and clean dataset, whereas on the other dataset with limited amount of data, traditional method outperforms all DL approaches. The achieved top-1 accuracy by the proposed model is 99.5% on 1-second voice instances from LibriSpeech dataset.

Self-segmentation of pass-phrase utterances for deep feature learning in text-dependent speaker verification

In this paper, we propose a novel method to segment and label pass-phrase utterances for training deep neural network (DNN) bottleneck (BN) features for text-dependent speaker verification (TD-SV). Specifically, gender-dependent hidden Markov models (HMMs) for monophones are first trained using the pass-phrase utterances that are disjoint from evaluation. Next, the trained HMMs are speaker-adapted and then used for segmenting and labeling these training utterances at the phone level. The resulted labeled data is subsequently used for training DNN models to discriminate gender-dependent phones for the purpose of extracting phone-discriminant BN features. This is in contrast to conventional approaches that apply a general-purpose, speaker-independent automatic speech recognition (ASR) system for generating segmentation and labels. The proposed method eliminates the need for a separate ASR system, which can additionally have the disadvantage of mismatch with the pass-phrase utterances in terms languages, dialects, domains, acoustic conditions and so on. Experiments are conducted on the RedDots challenge 2016 database of TD-SV using short utterances with Gaussian mixture model-universal background model and i-vector techniques. Experimental results demonstrate that the proposed method yields lower error rates in TD-SV when compared to a set of existing methods. A thorough ablation study further confirms the effectiveness of the method. Fusion in both score and feature levels also shows the complementary nature of the proposed features.

A GMM supervector approach for spoken Indian language identification for mismatch utterance length

Gaussian mixture model-universal background model (GMM UBM) supervectors are used to identify spoken Indian languages. The supervectors are calculated from short-time MFCC, its first and sec derivatives. The UBM builds a generalized Indian language model, and mean adaptation transforms it to a duration normalized language-specific GMM. Multi-class support vector machine and artificial neural network classifiers are used to identify language labels from the supervectors. Experimental evaluations are performed using 30 sec speech utterances from nine Indian languages comprised five Indo-Aryan and four Dravidian languages, extracted from all India radio broadcast news data-set. Eight smaller duration data-sets were manually derived to study the effect of training and test duration mismatch. In mismatch conditions, identification accuracy decreases with a decrease in test and train utterance duration. Investigations showed that the 32-mixture model with ANN classifier has optimal performance.

Combined i-Vector and Extreme Learning Machine Approach for Robust Speaker Identification and Evaluation with SITW 2016, NIST 2008, TIMIT Databases

In this article, a novel combined i-vector and an Extreme Learning Machine (ELM) is proposed for speaker identification. The ELM is chosen because it is fast to train and has a universal approximator property. Four combinations of features based on Mel Frequency Cepstral Coefficient and Power Normalized Cepstral Coefficient are used. Besides, seven fusion methods are exploited. The system is evaluated with three different databases, namely: the SITW 2006, NIST 2008, and the TIMIT database. This work employs the 2016 SITW database for the first time for speaker identification using the integration between the ELM and i-vector approach. From each database, 120 speakers with 1200 speech utterances are used (overall 360 speakers with 3600 speech utterances). Furthermore, comprehensive evaluations are exploited with a wide range of realistic background noise types (Stationary noise AWGN and Non-Stationary Noise types) with the handset effect. The proposed system is compared with the Gaussian Mixture Model-Universal Background Model (GMM-UBM) and other states of the art approaches. The results show that the i-vector method outperforms the GMM-UBM approach and other state- of-the-art methods under specific conditions, and that fusion techniques can be used to improve robustness to noise and handset effects.

Vocal Tract Length Perturbation for Text-Dependent Speaker Verification With Autoregressive Prediction Coding

In this letter, we propose a vocal tract length (VTL) perturbation method for text-dependent speaker verification (TD-SV), in which a set of TD-SV systems are trained, one for each VTL factor, and score-level fusion is applied to make a final decision. Next, we explore the bottleneck (BN) feature extracted by training deep neural networks with a self-supervised objective, autoregressive predictive coding (APC), for TD-SV and compare it with the well-studied speaker-discriminant BN feature. The proposed VTL method is then applied to APC and speaker-discriminant BN features. In the end, we combine the VTL perturbation systems trained on MFCC and the two BN features in the score domain. Experiments are performed on the RedDots challenge 2016 database of TD-SV using short utterances with Gaussian mixture model-universal background model and i-vector techniques. Results show the proposed methods significantly outperform the baselines.

Based on machine learning scheme to develop a smart robot embedded with GMM-UBM

Smart Robot embedded with GMM-UBM (Gaussian mixture model- universal background model) based on the machine learning scheme is presented in the article. Authors have designed a smart robot for the farmer and which is designed controlled by the concept of machine learning. On the other hand, the techniques of machine learning are applied to develop a smart robot for helping farmers recognize the environment conditions, e.g. weather, and disease protection in rice or plant. The smart robot is implemented to detect and to recognize the environment conditions around a fixed area. The sensing way through vision devices, such as camera, look like a human’s eye to distinguish various types of target. The QR code is deployed to simulate working conditions allows the robot to separate conditions and act according to conditions precisely. Besides, the smart robot is embedded with GMM-UBM algorithm for promoting the accuracy of recognition from the deployment of machine learning. The smart robot, mainly combines with AI (Artificial intelligence) techniques, consists of the following equipments: 1) a control movement subsystem, 2) a sensor control subsystem, and 3) an analysis subsystem. The researcher has determined the condition of the message options via QR code. In addition, the contents of the QR code tag will be processed a text message and saved to a memory device, once the reading is finished. The data analysis subsystem then reads the text and recommends the robot to move according to the specified conditions. The results from QR code data allow the smart robot to accurately collect many kinds of prefer data (e.g., climate data) in the farm at the specified location.

Two-Level Classification in Determining the Age and Gender Group of a Speaker

In this study, the classification of the speakers according to age and gender was discussed. Age and gender classes were first examined separately, and then by combining these classes a classification with a total of 7 classes was made. Speech signals represented by Mel-Frequency Cepstral Coefficients (MFCC) and delta parameters were converted into Gaussian Mixture Model (GMM) mean supervectors and classified with a Support Vector Machine (SVM). While the GMM mean supervectors were formed according to the Maximum-a-posteriori (MAP) adaptive GMM-Universal Background Model (UBM) configuration, the number of components was changed from 16 to 512, and the optimum number of components was decided. Gender classification accuracy of the system developed using aGender dataset was measured as 99.02% for two classes and 92.58% for three classes and age group classification accuracy was measured as 67.03% for female and 63.79% for male. In the classification of age and gender classes together in one step, an accuracy of 61.46% was obtained. In the study, a two-level approach was proposed for classifying age and gender classes together. According to this approach, the speakers were first divided into three classes as child, male and female, then males and females were classified according to their age groups and thus a 7-class classification was realized. This two-level approach was increased the accuracy of the classification in all other cases except when 32-component GMMs were used. While the highest improvement of 2.45% was achieved with 64 component GMMs, an improvement of 0.79 was achieved with 256 component GMMs.

Middle Eastern and North African English Speech Corpus (MENAESC): Automatic Identification of MENA English Accents

This study aims to explore the English accents in the Arab world. Although there are limited resources for a speech corpus that attempts to automatically identify the degree of accent patterns of an Arabic speaker of English, there is no speech corpus specialized for Arabic speakers of English in the Middle East and North Africa (MENA). To that end, different samples were collected in order to create the linguistic resource that we called Middle Eastern and North African English Speech Corpus (MENAESC). In addition to the “accent approach” applied in the field of automatic language/dialect recognition; we applied also the “macro-accent approach” -by employing Mel-Frequency Cepstral Coefficients (MFCC), Energy and Shifted Delta Cepstra (SDC) features and Gaussian Mixture Model-Universal Background Model (GMM-UBM) classifier- on four accents (Egyptian, Qatari, Syrian, and Tunisian accents) among the eleven accents that were selected based on their high population density in the location where the experiments were carried out. By using the Equal Error Rate percentage (EER%) for the assessment of our system effectiveness in the identification of MENA English accents using the two approaches mentioned above through the employ of the MENAESC, results showed we reached 1.5 to 2%, for “accent approach” and 2 to 3.5% for “macro-accents approach” for identification of MENA English. It also exhibited that the Qatari accent, of the 4 accents included, scored the lowest EER% for all tests performed. Taken together, the system effectiveness is not only affected by the approaches used, but also by the database size MENAESC and its characteristics. Moreover, it is impacted by the proficiency of the Arabic speakers of English and the influence of their mother tongue

Forensic speaker recognition: A new method based on extracting accent and language information from short utterances

This paper presents a new method for Forensic Speaker Recognition (FSR). The new method is based on extracting accent and language information from short utterances. Accent Classification (AC) and Language Identification (LI) play important role in the identification of people of different groups, communities and origins due to different speaking styles and native languages. In a multilingual society, the forensic experts use AC and LI to reduce search space for suspect recognition to regional and ethnic groups. In this paper, we use different baseline and deep learning methods to automate this process. The baseline methods used are Gaussian Mixture Model-Universal Background Model (GMM-UBM), i-vector and Gaussian Mixture Model-Support Vector Machine (GMM-SVM). The Mel-Frequency Cepstral Coefficients (MFCC) are used as speech features in the baseline methods. The deep learning methods used are Convolutional Neural Network (CNN) and Deep Neural Network (DNN). The recently proposed CNN based methods like VGGVox and GMM-CNN are used. VGGVox and GMM-CNN use speech spectrograms. In case of DNN, x-vectors method is used, which is based on DNN embedding. The experimental results show that GMM-SVM demonstrates better FSR performance compared to GMM-UBM and i-vector methods. Whereas, x-vectors method performs better than GMM-CNN and VGGVox methods. It also performs better than GMM-SVM method. The experimental results show that x-vectors method demonstrates 80.4% FSR accuracy. With AC, it achieves 85.4% accuracy. With LI, its accuracy is 90.2%. Whereas by combining AC and LI it obtains 95.1% accuracy. This shows that the proposed method based on AC and LI gives promising results.

Overlapping region reconstruction in nuclei image segmentation

Automatic screening systems play an increasingly important role in the diagnosis of pathologists. Image measurement and classification are the key techniques of automatic screening systems, which directly determine the performance. The distortion in grey and texture after overlapping nuclei segmentation seriously degrades the DNA content measurement and nuclei classification. In order to solve this problem, this paper presents a new method to reconstruct the pixels in overlapping regions based on the GMM-UBM (Gaussian mixture model–universal background model). In this method, a large amount of data are first used to train a GMM (named UBM). Then, the GMM of each nucleus is derived by maximizing a posteriori adaptation with the UBM and the normal grey value of this nucleus. The grey values are randomly generated by the GMM and filled to the overlapping region, with the offset to fine-tuning the Gaussian components. Finally, the image inpainting algorithm is used to repair the connected region. Experimental results show that this method can effectively recover the nucleus features, such as texture, grey and optical density, and improve the accuracy of nucleus measurement and classification.

Robust speaker recognition based on biologically inspired features

This paper proposes two speech parameterisation techniques for noise-robust speaker recognition: the normalised gammachirp cepstral coefficients (NGCC) and the perceptual linear predictive normalised gammachirp (PLPnGc). These techniques employ a biologically inspired auditory model that simulates the cochlea spectral behaviour. In an automatic speaker recognition (ASR) system, we consider the Gaussian mixture model-universal background model (GMM-UBM) for speaker modelling. The performances are evaluated in clean and noisy environments using Timit, Aurora, and Demand databases. The experimental results in noisy environments showed that the biologically inspired feature extraction techniques give a better recognition rate than state-of-the-art methods.