Reduction In Equal Error Rate Research Articles

Improving Short Utterance Speaker Recognition by Modeling Speech Unit Classes

Short utterance speaker recognition (SUSR) is highly challenging due to the limited enrollment and/or test data. We argue that the difficulty can be largely attributed to the mismatched prior distributions of the speech data used to train the universal background model (UBM) and those for enrollment and test. This paper presents a novel solution that distributes speech signals into a multitude of acoustic subregions that are defined by speech units, and models speakers within the subregions. To avoid data sparsity, a data-driven approach is proposed to cluster speech units into speech unit classes, based on which robust subregion models can be constructed. Further more, we propose a model synthesis approach based on maximum likelihood linear regression (MLLR) to deal with no-data speech unit classes. The experiments were conducted on a publicly available database SUD12. The results demonstrated that on a text-independent speaker recognition task where the test utterances are no longer than 2 seconds and mostly shorter than 0.5 seconds, the proposed sub-region modeling offered a 21.51% relative reduction in equal error rate (EER), compared with the standard GMM-UBM baseline. In addition, with the model synthesis approach, the performance can be greatly improved in scenarios where no enrollment data are available for some speech unit classes.

Wise teachers train better DNN acoustic models

Automatic speech recognition is becoming more ubiquitous as recognition performance improves, capable devices increase in number, and areas of new application open up. Neural network acoustic models that can utilize speaker-adaptive features, have deep and wide layers, or more computationally expensive architectures, for example, often obtain best recognition accuracy but may not be suitable for the given budget of computational and storage resources or latency required by the deployed system. We explore a straightforward training approach which takes advantage of highly accurate but expensive-to-evaluate neural network acoustic models by using their outputs to relabel training examples for easier-to-deploy models. Experiments on a large vocabulary continuous speech recognition task offer relative reductions in word error rate of up to 16.7 % over training with the hard aligned labels by effectively making use of large amounts of additional untranscribed data. Somewhat remarkably, the approach works well even when only two output classes are present. Experiments on a voice activity detection task give relative reductions in equal error rate of up to 11.5 % when using a convolutional neural network to relabel training examples for a feedforward neural network. An investigation into the hidden layer weight matrices finds that soft target-trained networks tend to produce weight matrices having fuller rank and slower decay in singular values than their hard target-trained counterparts, suggesting that more of the network's capacity is utilized for learning additional information giving better accuracy.

Rapid Language Identification

A critical challenge to automatic language identification (LID) is achieving accurate performance with the shortest possible speech segment in a rapid fashion. The accuracy to correctly identify the spoken language is highly sensitive to the duration of speech and is bounded by the amount of information available. The proposed approach for rapid language identification transforms the utterances to a low dimensional i-vector representation upon which language classification methods are applied. In order to meet the challenges involved in rapidly making reliable decisions about the spoken language, a highly accurate and computationally efficient framework of i-vector extraction is proposed. The LID framework integrates the approach of universal background model (UBM) fused total variability modeling. UBM-fused modeling yields the estimation of a more discriminant, single i-vector space. This way, it is also a computationally more efficient alternative than system level fusion. A further reduction in equal error rate is achieved by training the i-vector model on long duration speech utterances and by the deployment of a robust feature extraction scheme that aims to capture the relevant language cues under various acoustic conditions. Evaluation results on the DARPA RATS data corpus suggest the potential of performing successful automated language identification at the level of one second of speech or even shorter duration.

Speaker verification based on the fusion of speech acoustics and inverted articulatory signals

We propose a practical, feature-level and score-level fusion approach by combining acoustic and estimated articulatory information for both text independent and text dependent speaker verification. From a practical point of view, we study how to improve speaker verification performance by combining dynamic articulatory information with the conventional acoustic features. On text independent speaker verification, we find that concatenating articulatory features obtained from measured speech production data with conventional Mel-frequency cepstral coefficients (MFCCs) improves the performance dramatically. However, since directly measuring articulatory data is not feasible in many real world applications, we also experiment with estimated articulatory features obtained through acoustic-to-articulatory inversion. We explore both feature level and score level fusion methods and find that the overall system performance is significantly enhanced even with estimated articulatory features. Such a performance boost could be due to the inter-speaker variation information embedded in the estimated articulatory features. Since the dynamics of articulation contain important information, we included inverted articulatory trajectories in text dependent speaker verification. We demonstrate that the articulatory constraints introduced by inverted articulatory features help to reject wrong password trials and improve the performance after score level fusion. We evaluate the proposed methods on the X-ray Microbeam database and the RSR 2015 database, respectively, for the aforementioned two tasks. Experimental results show that we achieve more than 15% relative equal error rate reduction for both speaker verification tasks.

A perceptually-motivated low-complexity instantaneous linear channel normalization technique applied to speaker verification

This paper proposes a new set of speech features called Locally-Normalized Cepstral Coefficients (LNCC) that are based on Seneff's Generalized Synchrony Detector (GSD). First, an analysis of the GSD frequency response is provided to show that it generates spurious peaks at harmonics of the detected frequency. Then, the GSD frequency response is modeled as a quotient of two filters centered at the detected frequency. The numerator is a triangular band pass filter centered around a particular frequency similar to the ordinary Mel filters. The denominator term is a filter that responds maximally to frequency components on either side of the numerator filter. As a result, a local normalization is performed without the spurious peaks of the original GSD. Speaker verification results demonstrate that the proposed LNCC features are of low computational complexity and far more effectively compensate for spectral tilt than ordinary MFCC coefficients. LNCC features do not require the computation and storage of a moving average of the feature values, and they provide relative reductions in Equal Error Rate (EER) as high as 47.7%, 34.0% or 25.8% when compared with MFCC, MFCC+CMN, or MFCC+RASTA in one case of variable spectral tilt, respectively.

Optimization of the parameters characterizing sigmoidal rate-level functions based on acoustic features

This paper describes the development of an optimal sigmoidal rate-level function that is a component of many models of the peripheral auditory system. The optimization makes use of a set of criteria defined exclusively on the basis of physical attributes of the input sound that are inspired by physiological evidence. The criteria developed attempt to discriminate between a degraded speech signal and noise to preserve the maximum amount of information in the linear region of the sigmoidal curve, and to minimize the effects of distortion in the saturating regions. The performance of the proposed optimal sigmoidal function is validated by text-independent speaker-verification experiments with signals corrupted by additive noise at different SNRs. The experimental results suggest that the approach presented in combination with cepstral variance normalization can lead to relative reductions in equal error rate as great as 40% when compared with the use of baseline MFCC coefficients for some SNRs.

Combining log-spectral mean subtraction at different frequency resolutions for handset-channel compensation in single utterance speaker verification

Cepstral mean subtraction (CMS) is a well-known feature domain channel compensation technique employed to eliminate the effects of convolutive channel distortion. However, as the authors use in log-spectral mean subtraction (LSMS), the compensation might be applied in spectral domain before the filter-bank analysis with a higher-frequency resolution. LSMS can also be combined with CMS to further improve the recognition performance. In this study, the authors compare the performances of LSMS and CMS methods using a multi-channel, text-dependent single utterance speaker recognition database. In the experiments, the authors observe that LSMS outperforms CMS especially in the high false acceptance region. Moreover, the accuracy is further improved when the methods are combined together. With the combination, the authors achieve 15.5% relative reduction in equal error rate for no score normalisation and 9.4% for test normalisation cases when compared with the baseline CMS experiment.

Telephone Channel Compensation in Speaker Verification Using a Polynomial Approximation in the Log-Filter-Bank Energy Domain

This correspondence presents a novel feature-space channel compensation technique that models the convolutional distortion in the log-energy mel-filter domain by means of a polynomial approximation. The proposed parametric distortion model generates appropriate constraints in the spectral domain that help to improve the channel cancelling estimation with limited data. In a text-dependent speaker verification task, the polynomial-based channel estimation scheme can lead to reductions in equal error rate (EER) as great as 22% and 8% when compared with the baseline system and with the standard cepstral bias removal approach, respectively, with no significant increase in computational load.

Calibration of Confidence Measures in Speech Recognition

Most speech recognition applications in use today rely heavily on confidence measure for making optimal decisions. In this paper, we aim to answer the question: what can be done to improve the quality of confidence measure if we cannot modify the speech recognition engine? The answer provided in this paper is a post-processing step called confidence calibration, which can be viewed as a special adaptation technique applied to confidence measure. Three confidence calibration methods have been developed in this work: the maximum entropy model with distribution constraints, the artificial neural network, and the deep belief network. We compare these approaches and demonstrate the importance of key features exploited: the generic confidence-score, the application-dependent word distribution, and the rule coverage ratio. We demonstrate the effectiveness of confidence calibration on a variety of tasks with significant normalized cross entropy increase and equal error rate reduction.

On-Line Linear Combination of Classifiers Based on Incremental Information in Speaker Verification

A novel multiclassifier system (MCS) strategy is proposed and applied to a text-dependent speaker verification task. The presented scheme optimizes the linear combination of classifiers on an on-line basis. In contrast to ordinary MCS approaches, neither a priori distributions nor pre-tuned parameters are required. The idea is to improve the most accurate classifier by making use of the incremental information provided by the second classifier. The on-line multiclassifier optimization approach is applicable to any pattern recognition problem. The proposed method needs neither a priori distributions nor pre-estimated weights, and does not make use of any consideration about training/testing matching conditions. Results with Yoho database show that the presented approach can lead to reductions in equal error rate as high as 28%, when compared with the most accurate classifier, and 11% against a standard method for the optimization of linear combination of classifiers.

Utterance Verification Using State-Level Log-Likelihood Ratio with Frame and State Selection

This paper suggests utterance verification system using state-level log-likelihood ratio with frame and state selection. We use hidden Markov models for speech recognition and utterance verification as acoustic models and anti-phone models. The hidden Markov models have three states and each state represents different characteristics of a phone. Thus we propose an algorithm to compute state-level log-likelihood ratio and give weights on states for obtaining more reliable confidence measure of recognized phones. Additionally, we propose a frame selection algorithm to compute confidence measure on frames including proper speech in the input speech. In general, phone segmentation information obtained from speaker-independent speech recognition system is not accurate because triphone-based acoustic models are difficult to effectively train for covering diverse pronunciation and coarticulation effect. So, it is more difficult to find the right matched states when obtaining state segmentation information. A state selection algorithm is suggested for finding valid states. The proposed method using state-level log-likelihood ratio with frame and state selection shows that the relative reduction in equal error rate is 18.1% compared to the baseline system using simple phone-level log-likelihood ratios.

Online Signature Verification and Recognition: An Approach Based on Symbolic Representation

In this paper, we propose a new method of representing on-line signatures by interval valued symbolic features. Global features of on-line signatures are used to form an interval valued feature vectors. Methods for signature verification and recognition based on the symbolic representation are also proposed. We exploit the notions of writer dependent threshold and introduce the concept of feature dependent threshold to achieve a significant reduction in equal error rate. Several experiments are conducted to demonstrate the ability of the proposed scheme in discriminating the genuine signatures from the forgeries. We investigate the feasibility of the proposed representation scheme for signature verification and also signature recognition using all 16500 signatures from 330 individuals of the MCYT bimodal biometric database. Further, extensive experimentations are conducted to evaluate the performance of the proposed methods by projecting features onto Eigenspace and Fisherspace. Unlike other existing signature verification methods, the proposed method is simple and efficient. The results of the experimentations reveal that the proposed scheme outperforms several other existing verification methods including the state-of-the-art method for signature verification.

Utterance Verification Using Word Voiceprint Models Based on Probabilistic Distributions of Phone-Level Log-Likelihood Ratio and Phone Duration

This paper suggests word voiceprint models to verify the recognition results obtained from a speech recognition system. Word voiceprint models have word-dependent information based on the distributions of phone-level log-likelihood ratio and duration. Thus, we can obtain a more reliable confidence score for a recognized word by using its word voiceprint models that represent the more proper characteristics of utterance verification for the word. Additionally, when obtaining a log-likelihood ratio-based word voiceprint score, this paper proposes a new log-scale normalization function using the distribution of the phone-level log-likelihood ratio, instead of the sigmoid function widely used in obtaining a phone-level log-likelihood ratio. This function plays a role of emphasizing a mis-recognized phone in a word. This individual information of a word is used to help achieve a more discriminative score against out-of-vocabulary words. The proposed method requires additional memory, but it shows that the relative reduction in equal error rate is 16.9% compared to the baseline system using simple phone log-likelihood ratios.

Speaker recognition based on multi-subsystems likelihood scores fusion

To describe an approach of speaker clustering based on multi-subsystems likelihood scores fusion to the text-independent speaker recognition system with short speech data in various telephone microphone channels. The registered speakers were aggregated into clusters with 2 types of speaker model similarity measures, namely, Kullback-Leibler Divergence (KLD) and Generalized Likelihood Ratio (GLR). A single-layer perception network was built for each cluster, fusing the likelihood scores of 3 sub-systems with the speaker features of MFCC, LPCC and SSFE, respectively. Concerning the robustness of SSFE system and recognition accuracy of the other 2 systems, the 3 sub-systems complement each other with the fusion network in each cluster. Experimental results on NIST SRE 05's database show a relative equal error rate reduction of 10.3% and 8.7%, with KLD and GLR, respectively, with respect to an all-speaker-shared fusion network.

Out-of-Domain Utterance Detection Using Classification Confidences of Multiple Topics

One significant problem for spoken language systems is how to cope with users' out-of-domain (OOD) utterances which cannot be handled by the back-end application system. In this paper, we propose a novel OOD detection framework, which makes use of the classification confidence scores of multiple topics and applies a linear discriminant model to perform in-domain verification. The verification model is trained using a combination of deleted interpolation of the in-domain data and minimum-classification-error training, and does not require actual OOD data during the training process, thus realizing high portability. When applied to the phrasebook system, a single utterance read-style speech task, the proposed approach achieves an absolute reduction in OOD detection errors of up to 8.1 points (40% relative) compared to a baseline method based on the maximum topic classification score. Furthermore, the proposed approach realizes comparable performance to an equivalent system trained on both in-domain and OOD data, while requiring no OOD data during training. We also apply this framework to the machine-aided-dialogue corpus, a spontaneous dialogue speech task, and extend the framework in two manners. First, we introduce topic clustering which enables reliable topic confidence scores to be generated even for indistinct utterances, and second, we implement methods to effectively incorporate dialogue context. Integration of these two methods into the proposed framework significantly improves OOD detection performance, achieving a further reduction in equal error rate (EER) of 7.9 points

A dynamic in-search data selection method with its applications to acoustic modeling and utterance verification

In this paper, we propose a dynamic in-search data selection method to diagnose competing information automatically from speech data. In our method, the Viterbi beam search is used to decode all training data. During decoding, all partial paths within the beam are examined to identify the so-called competing-token and true-token sets for each individual hidden Markov model (HMM). In this work, the collected data tokens are used for acoustic modeling and utterance verification as two specific examples. In acoustic modeling, the true-token sets are used to adapt HMMs with a sequential maximum a posteriori adaptation method, while a generalized probabilistic descent-based discriminative training method is proposed to improve HMMs based on competing-token sets. In utterance verification, under the framework of likelihood ratio testing, the true-token sets are employed to train positive models for the null hypothesis and the competing-token sets are used to estimate negative models for the alternative hypothesis. All the proposed methods are evaluated in Bell Laboratories communicator system. Experimental results show that the new acoustic modeling method can consistently improve recognition performance over our best maximum likelihood estimation models, roughly 1% absolute reduction in word error rate. The results also show the new verification models can significantly improve the performance of utterance verification over the conventional anti models, almost relatively 30% reduction of equal error rate when identifying misrecognized words from the recognition results.

A new approach to utterance verification based on neighborhood information in model space

We propose to use neighborhood information in model space to perform utterance verification (UV). At first, we present a nested-neighborhood structure for each underlying model in model space and assume the underlying model's competing models sit in one of these neighborhoods, which is used to model alternative hypothesis in UV. Bayes factors (BF) is first introduced to UV and used as a major tool to calculate confidence measures based on the above idea. Experimental results in the Bell Labs communicator system show that the new method has dramatically improved verification performance when verifying correct words against mis-recognized words in the recognizer's output, relatively more than 20% reduction in equal error rate (EER) when comparing with the standard approach based on likelihood ratio testing and anti-models.

Efficient text-independent speaker verification with structural gaussian mixture models and neural network

We present an integrated system with structural Gaussian mixture models (SGMMs) and a neural network for purposes of achieving both computational efficiency and high accuracy in text-independent speaker verification. A structural background model (SBM) is constructed first by hierarchically clustering all Gaussian mixture components in a universal background model (UBM). In this way the acoustic space is partitioned into multiple regions in different levels of resolution. For each target speaker, a SGMM can be generated through multilevel maximum a posteriori (MAP) adaptation from the SBM. During test, only a small subset of Gaussian mixture components are scored for each feature vector in order to reduce the computational cost significantly. Furthermore, the scores obtained in different layers of the tree-structured models are combined via a neural network for final decision. Different configurations are compared in the experiments conducted on the telephony speech data used in the NIST speaker verification evaluation. The experimental results show that computational reduction by a factor of 17 can be achieved with 5% relative reduction in equal error rate (EER) compared with the baseline. The SGMM-SBM also shows some advantages over the recently proposed hash GMM, including higher speed and better verification performance.

Discriminating observation probability (DOP) HMM for speaker verification

This paper describes the use of a multiple codebook semi-continuous hidden Markov model (SCHMM) automatic speaker verification (ASV) system, which uses a novel technique for discriminative hidden Markov modelling known as discriminative observation probabilities (DOP). DOP is not a discriminative training technique, it is a method of constructing what is effectively a discriminating model by contrasting two standard HMMs so as to improve discrimination between the classes that those models represent. This paper experimentally evaluates the use of DOP HMMs for ASV. The experimental evaluation is based on a text-dependent task using isolated digits. The database contains 24 true (client) speakers and 100 casual impostors, recorded over the public telephone network in the United Kingdom. For speaker verification, DOP has similarities to the so-called speaker normalisation technique but it is a more flexible approach and experimental results are presented here which show it to be superior, even in its simplest form. An argument is also made for preferring DOP models to discriminatively trained models although this is not evaluated experimentally. The DOP technique can easily be added to an existing HMM system and requires no additional training. The DOP concept can potentially be applied to other applications of HMMs to two-class classification problems. An equal error rate (EER) of 0.21% (using speaker specific thresholds, on a sequence of 12 isolated digits) was obtained using multiple codebook DOP models. This represents a reduction in EER of 38% compared to a common speaker normalisation technique.