Universal Background Model Research Articles

Text-Independent Speaker Recognition in Clean and Noisy Backgrounds Using Modified VQ-LBG Algorithm

Speaker recognition is the process of identifying the proper speaker by analyzing the spectral shape of the speech signal. This process is done by extracting the desired features and matching the features of the speech signal. In this paper, we adopted the Mel frequency cepstrum coefficient (MFCC) technique for extracting the features from the speaker speech sample. These cepstrum coefficients are named as extracted features. The extracted MFCC features are given as input to the modified vector quantization via Linde–Buzo–Gray (modified VQ-LBG) process and expectation maximization (EM) algorithm. Vector quantization technique is mainly used for feature matching where a separate codebook will be generated for each speaker. The EM algorithm is utilized to develop the Gaussian mixture model–universal background model (GMM–UBM). In GMM–UBM model, k means cluster is summed up to consolidate data about the covariance structure of the information and the focuses of the inert Gaussians. From our analysis, the modified VQ-LBG algorithm gives better performance compared to the GMM–UBM model.

Robust Binaural Localization of a Target Sound Source by Combining Spectral Source Models and Deep Neural Networks

Despite there being clear evidence for top-down (e.g., attentional) effects in biological spatial hearing, relatively few machine hearing systems exploit top-down model-based knowledge in sound localisation. This paper addresses this issue by proposing a novel framework for binaural sound localisation that combines model-based information about the spectral characteristics of sound sources and deep neural networks (DNNs). A target source model and a background source model are first estimated during a training phase using spectral features extracted from sound signals in isolation. When the identity of the background source is not available, a universal background model can be used. During testing, the source models are used jointly to explain the mixed observations and improve the localisation process by selectively weighting source azimuth posteriors output by a DNN-based localisation system. To address the possible mismatch between training and testing, a model adaptation process is further employed on-the-fly during testing, which adapts the background model parameters directly from the noisy observations in an iterative manner. The proposed system therefore combines model-based and data-driven information flow within a single computational framework. The evaluation task involved localisation of a target speech source in the presence of an interfering source and room reverberation. Our experiments show that by exploiting model-based information in this way, sound localisation performance can be improved substantially under various noisy and reverberant conditions.

Speaker identification using multi-modal i-vector approach for varying length speech in voice interactive systems

The development in the interface of smart devices has lead to voice interactive systems. An additional step in this direction is to enable the devices to recognize the speaker. But this is a challenging task because the interaction involves short duration speech utterances. The traditional Gaussian mixture models (GMM) based systems have achieved satisfactory results for speaker recognition only when the speech lengths are sufficiently long. The current state-of-the-art method utilizes i-vector based approach using a GMM based universal background model (GMM-UBM). It prepares an i-vector speaker model from a speaker’s enrollment data and uses it to recognize any new test speech. In this work, we propose a multi-model i-vector system for short speech lengths. We use an open database THUYG-20 for the analysis and development of short speech speaker verification and identification system. By using an optimum set of mel-frequency cepstrum coefficients (MFCC) based features we are able to achieve an equal error rate (EER) of 3.21% as compared to the previous benchmark score of EER 4.01% on the THUYG-20 database. Experiments are conducted for speech lengths as short as 0.25 s and the results are presented. The proposed method shows improvement as compared to the current i-vector based approach for shorter speech lengths. We are able to achieve improvement of around 28% even for 0.25 s speech samples. We also prepared and tested the proposed approach on our own database with 2500 speech recordings in English language consisting of actual short speech commands used in any voice interactive system.

Deep neural network based i-vector mapping for speaker verification using short utterances

Text-independent speaker recognition using short utterances is a highly challenging task due to the large variation and content mismatch between short utterances. I-vector and probabilistic linear discriminant analysis (PLDA) based systems have become the standard in speaker verification applications, but they are less effective with short utterances. In this paper, we first compare two state-of-the-art universal background model (UBM) training methods for i-vector modeling using full-length and short utterance evaluation tasks. The two methods are Gaussian mixture model (GMM) based (denoted I-vector_GMM) and deep neural network (DNN) based (denoted as I-vector_DNN) methods. The results indicate that the I-vector_DNN system outperforms the I-vector_GMM system under various durations (from full length to 5 s). However, the performances of both systems degrade significantly as the duration of the utterances decreases. To address this issue, we propose two novel nonlinear mapping methods which train DNN models to map the i-vectors extracted from short utterances to their corresponding long-utterance i-vectors. The mapped i-vector can restore missing information and reduce the variance of the original short-utterance i-vectors. The proposed methods both model the joint representation of short and long utterance i-vectors: the first method trains an autoencoder first using concatenated short and long utterance i-vectors and then uses the pre-trained weights to initialize a supervised regression model from the short to long version; the second method jointly trains the supervised regression model with an autoencoder reconstructing the short utterance i-vector itself. Experimental results using the NIST SRE 2010 dataset show that both methods provide significant improvement and result in a 24.51% relative improvement in Equal Error Rates (EERs) from a baseline system. In order to learn a better joint representation, we further investigate the effect of a deep encoder with residual blocks, and the results indicate that the residual network can further improve the EERs of a baseline system by up to 26.47%. Moreover, in order to improve the short i-vector mapping to its long version, an additional vector, which represents the average value of phoneme posteriors across frames, is also added to the input, and results in a 28.43% improvement. When further testing the best-validated models of SRE10 on the Speaker In The Wild (SITW) dataset, the methods result in a 23.12% improvement on arbitrary-duration (1–5 s) short-utterance conditions.

A Pre-classification-Based Language Identification for Northeast Indian Languages Using Prosody and Spectral Features

This paper is aimed at developing a two-stage language identification (LID) system for Northeast Indian languages. In the first stage, languages are pre-classified into tonal and non-tonal categories, and in the second stage, individual languages are identified from languages of the corresponding category. In this work, new parameters to model the prosodic characteristics of the speech signal have been proposed for pre-classification as well as individual language identification. Also, the effectiveness of spectral features, namely Mel-frequency cepstral coefficient (MFCC) and their combination with prosodic features, has been studied for pre-classification task. The usefulness of MFCC with their delta and acceleration coefficients in combination with prosodic features has been investigated for individual language identification. The performance of the system is analyzed for the features extracted of different analysis units, such as syllable, disyllable, word, and utterance. Comparative performance analysis of three different classifiers, namely artificial neural network (ANN), Gaussian mixture model–Universal background model (GMM–UBM), and i-vector based support vector machine (i-vector based SVM), has been made for pre-classification as well as individual language identification. A new database, NIT Silchar language database (NITS-LD), has been developed for seven NE Indian languages using All India Radio broadcast news. The experimental analysis suggests that the parameters proposed to represent the prosodic characteristics help to improve the performance of both the stages and show improvements over existing parameters by as much as 7.4%, 11.9%, and 9.1% for 30 s, 10 s, and 3 s test data, respectively, in the pre-classification stage. Of the baseline single-stage systems, GMM–UBM provides the highest accuracies of 80%, 76.8%, and 72% for 30 s, 10 s, and 3 s test data, respectively. In the proposed system, the combination of the ANN model in pre-classification stage and the GMM–UBM model in individual language identification stage provides the highest accuracies, and it shows the improvements over the baseline system by 7.2%, 7%, and 4.9% for 30 s, 10 s, and 3 s test data. For OGI-Multilingual (OGI-MLTS) database, improvements of 8.1%, 7.4%, and 5.7% for 30 s, 10 s, and 3 s test data, respectively, are observed over the baseline LID system.

Automatic speaker, age-group and gender identification from children’s speech

A speech signal contains important paralinguistic information, such as the identity, age, gender, language, accent, and the emotional state of the speaker. Automatic recognition of these types of information in adults’ speech has received considerable attention, however there has been little work on children’s speech. This paper focuses on speaker, gender, and age-group recognition from children’s speech. The performances of several classification methods are compared, including Gaussian Mixture Model–Universal Background Model (GMM–UBM), GMM–Support Vector Machine (GMM–SVM) and i-vector based approaches. For speaker recognition, error rate decreases as age increases, as one might expect. However for gender and age-group recognition the effect of age is more complex due mainly to consequences of the onset of puberty. Finally, the utility of different frequency bands for speaker, age-group and gender recognition from children’s speech is assessed.

Study of MFCC and IHC Feature Extraction Methods With Probabilistic Acoustic Models for Speaker Biometric Applications

Voice is an important human trait in natural human-to-human interaction / communication for identifying a person. So voice can be regarded as a biometric measure for recognizing or identifying the person similar to other biometric measures such as face, iris and fingerprints. Speaker recognition is a class of voice recognition where speaker is identified from the speech rather than the message. Automatic speaker recognition (SR) is an approach to identify people based on features extracted from speech utterances. The major task in any speaker recognition is to extract useful features and allow meaningful patterns of speaker models. This paper compares the performance of two feature extraction techniques Mel Frequency Cepstral Coefficient (MFCC) and Inner Hair Cell Coefficient (IHC) with two different modelling methods Gaussian Mixture Model - Universal background model (GMM - UBM) and i- vector approach. In this experiment speech samples of 600 speakers from TIMIT database with 10 utterances of each speaker are taken for identifying the speaker. A text independent speaker recognition system was implemented and this study resulted in an inference of MFCC feature outperforms IHC feature for both GMM and i vector. The performance of voiced speech IHC feature which simulates the physiological behaviour of human ear is better in terms of accuracy than full speech (voiced and unvoiced) in GMM.

I-vector Extraction for Speaker Recognition Based on Dimensionality Reduction

In the domain of speaker recognition, many methods have been proposed over time. The technology for automatic speaker recognition has now reached a good level of performance but there is still need of improvement. In this paper, a new low-dimensional speaker- and channel-dependent space is defined using a simple factor analysis also known as i-vector. This space is named the total variability space because it models both speaker and channel variabilities. The i-vector subspace modelling is one of the recent methods that have become the state of the art technique in this domain. This method largely provides the benefit of modelling both the intra-domain and inter-domain variabilities into the same low dimensional space. In this study, 2656 syllables bio-acoustic signals from 55 species of frog taken from Intelligent Biometric Group, USM database are used for frog identification system. Parameters of the system are initially tuned such as Universal Background Model (UBM) size (32, 64 and 128 Gaussians) and i-vector dimensionality (100, 200 and 400 dimensions). To the end, we assess the effect of the parameter tuned and record the computation time. We observed that, the accuracy for smaller UBM size and higher i-vector dimensionality outperforms others with result of 91.11% is achieved. From this research, it can be concluded that UBM size and i-vector dimensionality effect the accuracy of frog identification based on i-vector.

Robust Voice Liveness Detection and Speaker Verification Using Throat Microphones

While having a wide range of applications, automatic speaker verification (ASV) systems are vulnerable to spoofing attacks, in particular, replay attacks that are effective and easy to implement. Most prior work on detecting replay attacks uses audio from a single acoustic microphone only, leading to difficulties in detecting high-end replay attacks close to indistinguishable from live human speech. In this paper, we study the use of a special body-conducted sensor, throat microphone (TM), for combined voice liveness detection (VLD) and ASV in order to improve both robustness and security of ASV against replay attacks. We first investigate the possibility and methods of attacking a TM-based ASV system, followed by a pilot data collection. Second, we study the use of spectral features for VLD using both single-channel and dual-channel ASV systems. We carry out speaker verification experiments using Gaussian mixture model with universal background model (GMM-UBM) and i-vector based systems on a dataset of 38 speakers collected by us. We have achieved considerable improvement in recognition accuracy, with the use of dual-microphone setup. In experiments with noisy test speech, the false acceptance rate (FAR) of the dual-microphone GMM-UBM based system for recorded speech reduces from 69.69% to 18.75%. The FAR of replay condition further drops to 0% when this dual-channel ASV system is integrated with the new dual-channel voice liveness detector.

Acoustic Feature Analysis and Discriminative Modeling for Language Identification of Closely Related South-Asian Languages

With the advancement in technology, communication between people around the world from different linguistic backgrounds is increasing gradually, resulting in the requirement of language identification services. Language identification techniques extract distinguishable information as features of a language from the speech corpora to differentiate one language from other. Without publicly available speech corpora, comparison between different techniques will not be much reliable. This paper investigates state-of-the-art features and techniques for language identification of under-resource and closely related languages, namely Pashto, Punjabi, Sindhi, and Urdu. For language identification, speech corpus is designed and collected for mentioned languages. The dataset is a read speech data collected over telephone network (mobile and landline) from different regions of Pakistan. The speech corpus is annotated at the sentence level using X-SAMPA, its orthographic transcription is also provided, and verified data are divided into training and evaluation sets. Mel-frequency cepstral coefficients and their shifted delta cepstral features are used to develop language identification system of target languages. Gaussian mixture model with universal background model (GMM-UBM)-based and I-vector-based language identification approaches are investigated. The results show that GMM-UBM is more effective than the I-vector for language identification of short duration test utterances.

Performance comparison of multitaper techniques for speaker verification with expressive speech

In this paper, we provide a comparative study of spectral front-end features used as representations for speech signals by processing multitaper magnitude and phase spectra, for speaker verification with expressive speech. In particular, the multitaper modified group delay function (MT-MOGDF) and multitaper magnitude (MT-MAG) spectra of the speech signals are employed to obtain low variance estimates of speech spectra. We observe that the cues that aid in representation of expressive speech are evident in the MT-MOGDF spectrum than the MT-MAG spectrum in terms of mean Formant value and Formant bandwidth. Our extensive experimental study on a speaker verification system with a Gaussian mixture model based universal background model classifier on expressive speech using the IITKGP-SESC and EMODB databases show that MT-MOGDF performs better than MT-MAG technique, in terms of equal error rate and minimum decision cost function. This improvement due to MT-MOGDF is owed to a better representation and a low-variance estimate of the speech spectrum. Our results highlight the utility of MT-MOGDF as a potential alternative for MT-MAG representation for speaker verification problems in general.

Deep neural network training for whispered speech recognition using small databases and generative model sampling

State-of-the-art speech recognition solutions currently employ hidden Markov models (HMMs) to capture the time variability in a speech signal and deep neural networks (DNNs) to model the HMM state distributions. It has been shown that DNN–HMM hybrid systems outperform traditional HMM and Gaussian mixture model (GMM) hybrids in many applications. This improvement is mainly attributed to the ability of DNNs to model more complex data structures. However, having sufficient data samples is one key point in training a high accuracy DNN as a discriminative model. This barrier makes DNNs unsuitable for many applications with limited amounts of data. In this study, we introduce a method to produce an excessive amount of pseudo-samples that requires availability of only a small amount of transcribed data from the target domain. In this method, a universal background model (UBM) is trained to capture a parametric estimate of the data distributions. Next, random sampling is used to generate a large amount of pseudo-samples from the UBM. Frame-Shuffling is then applied to smooth the temporal cepstral trajectories in the generated pseudo-sample sequences to better resemble the temporal characteristics of a natural speech signal. Finally, the pseudo-sample sequences are combined with the original training data to train the DNN–HMM acoustic model of a speech recognizer. The proposed method is evaluated on small-sized sets of neutral and whisper data drawn from the UT-Vocal Effort II corpus. It is shown that phoneme error rates (PERs) of a DNN–HMM based speech recognizer are considerably reduced when incorporating the generated pseudo-samples in the training process, with + 79.0 and + 45.6% relative PER improvements for neutral–neutral training/test and whisper–whisper training/test scenarios, respectively.

Data selection for i-vector based automatic speaker verification anti-spoofing

State-of-the-art i-vector based automatic speaker verification (ASV) systems lead to considerably high performance and thus voice becomes one of the most important biometric modality for person authentication. However, similar to other biometrics, ASV systems are highly vulnerable to spoofing attacks. Therefore, developing countermeasures for detecting spoofing attacks plays an important role against the concerns regarding the reliability of ASV systems. Recent studies have shown that simple Gaussian mixture model (GMM) classifier outperforms i-vector countermeasures. In this study, we focus on improving the spoofing detection performance of i-vector system using cosine and probabilistic linear discriminant analysis (PLDA) scoring. Experimental results conducted on ASVspoof 2015 database reveals that the data used to train the two key elements of i-vector system, universal background model (UBM) and the i-vector extractor (T-matrix), play an important role on spoofing detection performance. In this paper, we study the effect of using different type of data (genuine/human or spoofed) to train these two elements and their performance on spoofing detection. In particular, extracting i-vectors using UBM trained with genuine (human) speech utterances and T-matrix trained from both genuine and spoofed utterances leads to 50% performance improvement on spoofing detection. With the proposed scheme, unlike the previous results, i-vector countermeasure outperforms GMM classifier. Finally, experimental results shows that recently proposed constant Q cepstral coefficients (CQCC) shows superior performance in comparison to standard Mel-frequency cepstral coefficients (MFCC).

Improving the Accuracy and Training Speed of Motor Imagery Brain-Computer Interfaces Using Wavelet-Based Combined Feature Vectors and Gaussian Mixture Model-Supervectors.

In this paper, we propose a set of wavelet-based combined feature vectors and a Gaussian mixture model (GMM)-supervector to enhance training speed and classification accuracy in motor imagery brain–computer interfaces. The proposed method is configured as follows: first, wavelet transforms are applied to extract the feature vectors for identification of motor imagery electroencephalography (EEG) and principal component analyses are used to reduce the dimensionality of the feature vectors and linearly combine them. Subsequently, the GMM universal background model is trained by the expectation–maximization (EM) algorithm to purify the training data and reduce its size. Finally, a purified and reduced GMM-supervector is used to train the support vector machine classifier. The performance of the proposed method was evaluated for three different motor imagery datasets in terms of accuracy, kappa, mutual information, and computation time, and compared with the state-of-the-art algorithms. The results from the study indicate that the proposed method achieves high accuracy with a small amount of training data compared with the state-of-the-art algorithms in motor imagery EEG classification.

Non‐parallel training for voice conversion using background‐based alignment of GMMs and INCA algorithm

Most of the voice conversion (VC) researches have used parallel training corpora to train the conversion function. However, in practice it is not always possible to gather parallel corpora, so the need for non-parallel training methods arises. As a successful non-parallel method, nearest neighbour search step and a conversion step alignment method (INCA) algorithm has attracted a lot of attention in recent years. In this study, the authors propose a new method of non-parallel VC which is based on the INCA algorithm. The authors' method effectively solves the initialisation problem of INCA algorithm. Their proposed initialisation for INCA is done with alignment of Gaussian mixture models (GMM) using universal background model. Results of objective and subjective experiments determined that the authors' proposed method improves the INCA algorithm. It is observed that this superiority holds for different sizes of training material from 10 to 50 training sentences. In terms of mean opinion score, the authors' method scores 0.25 higher in the case of quality and 0.2 higher in the case of similarity to the target speaker compared with traditional INCA. It seems that the authors' proposed method is a suitable frame alignment method for non-parallel corpora in VC task.

Gammatone filterbank and symbiotic combination of amplitude and phase-based spectra for robust speaker verification under noisy conditions and compression artifacts

The main novelty of this work resides in incorporating a Gammatone filter-bank as a substitute of the Mel filter-bank in the extraction pipeline of the Product Spectrum PS. The proposed feature is dubbed the Gammatone Product-Spectrum Cepstral coefficients GPSCC. Experimental results are undertaken on TIMIT and noisy TIMIT corpora using the Gaussian Mixture Model with Universal Background Model (GMM-UBM) recognition algorithm. Performance evaluations indicate that GPSCC shows a drastic reduction in Equal Error Rates compared to other related features and this gain in performance is more pronounced at low signal to noise ratios. Also, our study demonstrates the merit of the Gammatone filter-bank in improving robustness to codec-degraded speech at different bit rates. Furthermore, the proposed GPSCC feature achieves the best verification performance under aggressive compression. Interestingly, at 6.60 kbps we observe that GPSCC achieves an absolute error reduction of 12% compared to the Mel Frequency Cepstral Coefficients (MFCC).

Towards a Continuous Biometric System Based on ECG Signals Acquired on the Steering Wheel.

Electrocardiogram signals acquired through a steering wheel could be the key to seamless, highly comfortable, and continuous human recognition in driving settings. This paper focuses on the enhancement of the unprecedented lesser quality of such signals, through the combination of Savitzky-Golay and moving average filters, followed by outlier detection and removal based on normalised cross-correlation and clustering, which was able to render ensemble heartbeats of significantly higher quality. Discrete Cosine Transform (DCT) and Haar transform features were extracted and fed to decision methods based on Support Vector Machines (SVM), k-Nearest Neighbours (kNN), Multilayer Perceptrons (MLP), and Gaussian Mixture Models - Universal Background Models (GMM-UBM) classifiers, for both identification and authentication tasks. Additional techniques of user-tuned authentication and past score weighting were also studied. The method’s performance was comparable to some of the best recent state-of-the-art methods (94.9% identification rate (IDR) and 2.66% authentication equal error rate (EER)), despite lesser results with scarce train data (70.9% IDR and 11.8% EER). It was concluded that the method was suitable for biometric recognition with driving electrocardiogram signals, and could, with future developments, be used on a continuous system in seamless and highly noisy settings.

Evaluation of MSR Identity Toolbox under conditions reflecting those of a real forensic case (forensic_eval_01)

This paper presents an evaluation of the Microsoft Research Identity Toolbox version 1.0 developed at Microsoft Research, as a tool for forensic voice comparison under conditions reflecting those of a real forensic case. For this purpose we implement two systems: the first is based on Gaussian mixture model - universal background model (GMM-UBM) and the second on i-vectors with probabilistic linear discriminant analysis (i-vector PLDA). Three different feature-level mismatch compensation techniques were tested, before and after the application of voice activity detection (VAD). The three techniques were global cepstral mean subtraction (CMS), global cepstral mean and variance normalization (CMVN), and local feature warping (FW).

Classification of ADHD and Non-ADHD Subjects Using a Universal Background Model.

ADHD affects a major portion of our children, predominantly boys. Upon diagnosis treatment can be offered that is usually quite effective. Diagnosis is generally based on subjective observation and interview. As a result, an objective test for the detection or presence of ADHD is considered very desirable. Based on EEG, across multiple channels, using autoregressive model parameters as features, ADHD detection is approached here in analogy with the imposter problem known from speaker verification. Gaussian mixture models are used to define ADHD and universal background models so that a likelihood ratio detector can be designed. The efficacy of this approach is reflected in the traditional detector performance measures of the area-under-the-curve and equal-error-probability. The results - based on a limited database of males, approximately 6 years of age - indicate that high probability of detection and low equal error rate can be achieved simultaneously with the proposed approach, when using EEG collected during an attention network task. The effect of using contaminated data is investigated as well.

Incorporating pass-phrase dependent background models for text-dependent speaker verification

In this paper, we propose pass-phrase dependent background models (PBMs) for text-dependent (TD) speaker verification (SV) to integrate the pass-phrase identification process into the conventional TD-SV system, where a PBM is derived from a text-independent background model through adaptation using the utterances of a particular pass-phrase. During training, pass-phrase specific target speaker models are derived from the particular PBM using the training data for the respective target model. While testing, the best PBM is first selected for the test utterance in the maximum likelihood (ML) sense and the selected PBM is then used for the log likelihood ratio (LLR) calculation with respect to the claimant model. The proposed method incorporates the pass-phrase identification step in the LLR calculation, which is not considered in conventional standalone TD-SV systems. The performance of the proposed method is compared to conventional text-independent background model based TD-SV systems using either Gaussian mixture model (GMM)-universal background model (UBM) or hidden Markov model (HMM)-UBM or i-vector paradigms. In addition, we consider two approaches to build PBMs: speaker-independent and speaker-dependent. We show that the proposed method significantly reduces the error rates of text-dependent speaker verification for the non-target types: target-wrong and impostor-wrong while it maintains comparable TD-SV performance when impostors speak a correct utterance with respect to the conventional system. Experiments are conducted on the RedDots challenge and the RSR2015 databases that consist of short utterances.