From single to multiple enrollment i-vectors: Practical PLDA scoring variants for speaker verification

The availability of multiple utterances (and hence, i-vectors) for speaker enrollment brings up several alternatives for their utilization with probabilistic linear discriminant analysis (PLDA). This paper provides an overview of their effective utilization, from a practical viewpoint. We derive expressions for the evaluation of the likelihood ratio for the multi-enrollment case, with details on the computation of the required matrix inversions and determinants. The performance of five different scoring methods, and the effect of i-vector length normalization is compared experimentally. We conclude that length normalization is a useful technique for all but one of the scoring methods considered, and averaging i-vectors is the most effective out of the methods compared. We also study the application of multicondition training on the PLDA model. Our experiments indicate that multicondition training is more effective in estimating PLDA hyperparameters than it is for likelihood computation. Finally, we look at the effect of the configuration of the enrollment data on PLDA scoring, studying the properties of conditional dependence and number-of-enrollment-utterances per target speaker. Our experiments indicate that these properties affect the performance of the PLDA model. These results further support the conclusion that i-vector averaging is a simple and effective way to process multiple enrollment utterances.     

基于T矩阵归一化PLDA的说话人确认

利用i-vector/PLDA模型进行说话人确认时,对于不定时间的语音,由于将长度归一化后的i-vector转化到PLDA模型时,伴随着不确定性的扭曲和缩放,影响识别率。本文通过对全变量空间矩阵T的列向量执行归一化,代替在PLDA模型上对i-vector进行长度归一化,避免因在i-vector上执行长度归一化,导致转移到PLDA模型上产生不良的扭曲。实验结果表明,该方法得到和长度归一化相似的效果,部分效果要优于长度归一化。     

概率线性判别分析在语音命令词置信度判决中的应用

置信度判决用于确定语音数据与模型之间的匹配程度,可以发现语音命令系统中的识别错误,提高其可靠性.近年来,基于身份矢量(identity vector,i-vector)以及概率线性判别分析(Probabilistic Linear Discriminant Analysis,PLDA)的方法在说话人识别任务中取得了显著效果.本文尝试将i-vector以及PLDA模型作为一种命令词识别结果置信度分析方法,其无需声学模型、语言模型支撑,且实验表明性能良好.在此基础上,针对i-vector在刻画时序信息方面的不足,尝试将该系统与DTW融合,有效提升了系统对音频时序的鉴别能力.     

Simplified supervised i-vector modeling with application to robust and efficient language identification and speaker verification

This paper presents a simplified and supervised i-vector modeling approach with applications to robust and efficient language identification and speaker verification. First, by concatenating the label vector and the linear regression matrix at the end of the mean supervector and the i-vector factor loading matrix, respectively, the traditional i-vectors are extended to label-regularized supervised i-vectors. These supervised i-vectors are optimized to not only reconstruct the mean supervectors well but also minimize the mean square error between the original and the reconstructed label vectors to make the supervised i-vectors become more discriminative in terms of the label information. Second, factor analysis (FA) is performed on the pre-normalized centered GMM first order statistics supervector to ensure each gaussian component's statistics sub-vector is treated equally in the FA, which reduces the computational cost by a factor of 25 in the simplified i-vector framework. Third, since the entire matrix inversion term in the simplified i-vector extraction only depends on one single variable (total frame number), we make a global table of the resulting matrices against the frame numbers’ log values. Using this lookup table, each utterance's simplified i-vector extraction is further sped up by a factor of 4 and suffers only a small quantization error. Finally, the simplified version of the supervised i-vector modeling is proposed to enhance both the robustness and efficiency. The proposed methods are evaluated on the DARPA RATS dev2 task, the NIST LRE 2007 general task and the NIST SRE 2010 female condition 5 task for noisy channel language identification, clean channel language identification and clean channel speaker verification, respectively. For language identification on the DARPA RATS, the simplified supervised i-vector modeling achieved 2%, 16%, and 7% relative equal error rate (EER) reduction on three different feature sets and sped up by a factor of more than 100 against the baseline i-vector method for the 120 s task. Similar results were observed on the NIST LRE 2007 30 s task with 7% relative average cost reduction. Results also show that the use of Gammatone frequency cepstral coefficients, Mel-frequency cepstral coefficients and spectro-temporal Gabor features in conjunction with shifted-delta-cepstral features improves the overall language identification performance significantly. For speaker verification, the proposed supervised i-vector approach outperforms the i-vector baseline by relatively 12% and 7% in terms of EER and norm old minDCF values, respectively.     

Combining evidences from excitation source and vocal tract system features for Indian language identification using deep neural networks

In this paper, a combination of excitation source information and vocal tract system information is explored for the task of language identification (LID). The excitation source information is represented by features extracted from linear prediction (LP) residual signal called the residual cepstral coefficients (RCC). Vocal tract system information is represented by the mel frequency cepstral coefficients (MFCC). In order to incorporate additional temporal information, shifted delta cepstra (SDC) are computed. An LID system is built using SDC over both MFCC and RCC features individually and evaluated based on their equal error rate (EER). Experiments have been performed on a dataset consisting of 13 Indian languages with about 115 h for training and 30 h for testing using a deep neural network (DNN), DNN with attention (DNN-WA) and a state-of-the-art i-vector system. DNN-WA outperforms the baseline i-vector system. An EER of 9.93 and 6.25% are achieved using RCC and MFCC features respectively. By combining evidence from both features using a late fusion mechanism, an EER of 5.76% is obtained. This result indicates the complementary nature of the excitation source information to that of the widely used vocal tract system information for the task of LID.     

基于DNN处理的鲁棒性I-Vector说话人识别算法

提出了一种将基于深度神经网络（Deep Neural Network,DNN）特征映射的回归分析模型应用到身份认证矢量（identity vector,i-vector）/概率线性判别分析（Probabilistic Linear Discriminant Analysis,PLDA）说话人系统模型中的方法。DNN通过拟合含噪语音和纯净语音i-vector之间的非线性函数关系,得到纯净语音i-vector的近似表征,达到降低噪声对系统性能影响的目的。在TIMIT数据集上的实验验证了该方法的可行性和有效性。     

基于角度间隔嵌入特征的端到端声纹识别模型

针对传统身份认证矢量（i-vector）与概率线性判别分析（PLDA）结合的声纹识别模型步骤繁琐、泛化能力较弱等问题，构建了一个基于角度间隔嵌入特征的端到端模型。该模型特别设计了一个深度卷积神经网络，从语音数据的声学特征中提取深度说话人嵌入；选择基于角度改进的A-Softmax作为损失函数，在角度空间中使模型学习到的不同类别特征始终存在角度间隔并且同类特征间聚集更紧密。在公开数据集VoxCeleb2上进行的测试表明，与i-vector结合PLDA的方法相比，该模型在说话人辨认中的Top-1和Top-5上准确率分别提高了58.9%和30%；而在说话人确认中的最小检测代价和等错误率上分别减小了47.9%和45.3%。实验结果验证了所设计的端到端模型更适合在多信道、大规模的语音数据集上学习到有类别区分性的特征。     

A study on the roles of total variability space and session variability modeling in speaker recognition

Speaker verification (SV) using i-vector concept becomes state-of-the-art. In this technique, speakers are projected onto the total variability space and represented by vectors called i-vectors. During testing, the i-vectors of the test speech segment and claimant are conditioned to compensate for the session variability before scoring. So, i-vector system can be viewed as two processing blocks: one is total variability space and the other is post-processing module. Several questions arise, such as, (i) which part of the i-vector system plays a major role in speaker verification: total variability space or post-processing task; (ii) is the post-processing module intrinsic to the total variability space? The motivation of this paper is to partially answer these questions by proposing several simpler speaker characterization systems for speaker verification, where speakers are represented by their speaker characterization vectors (SCVs). The SCVs are obtained by uniform segmentation of the speakers gaussian mixture models (GMMs)- and maximum likelihood linear regression (MLLR) super-vectors. We consider two adaptation approaches for GMM super-vector: one is maximum a posteriori and other is MLLR. Similarly to the i-vector, SCVs are post-processed for session variability compensation during testing. The proposed system shows promising performance when compared to the classical i-vector system which indicates that the post-processing task plays an major role in i-vector based SV system and is not intrinsic to the total variability space. All experimental results are shown on NIST 2008 SRE core condition.     

基于i-向量和PCA字典学习稀疏表示的说话人确认

稀疏表示以其出色的分类性能成为说话人确认研究的热点,其中过完备字典的构建是关键,直接影响其性能。为了提高说话人确认系统的鲁棒性,同时解决稀疏表示过完备字典中存在噪声及信道干扰信息的问题,提出一种基于i-向量的主成分稀疏表示字典学习算法。该算法在高斯通用背景模型的基础上提取说话人的i-向量,并使用类内协方差归一化技术对i-向量进行信道补偿;根据信道补偿后的说话人i-向量的均值向量估计其信道偏移空间,在该空间采用主成分分析方法提取低维信道偏移主分量,用于重新计算说话人i-向量,从而达到进一步抑制i-向量中信道干扰的目的;将新的i-向量作为字典原子构建高鲁棒性稀疏表示过完备字典。在测试阶段,测试语音的i-向量在该字典上寻找其稀疏表示系数向量,根据系数向量对测试i-向量的重构误差确定目标说话人。仿真实验表明,该算法具有良好的识别性能。     

A bilevel framework for joint optimization of session compensation and classification for speaker identification

The i-vector framework based system is one of the most popular systems in speaker identification (SID). In this system, session compensation is usually employed first and then the classifier. For any session-compensated representation of i-vector, there is a corresponding identification result, so that both the stages are related. However, in current SID systems, session compensation and classifier are usually optimized independently. An incomplete knowledge about the session compensation to the identification task may lead to involving uncertainties. In this paper, we propose a bilevel framework to jointly optimize session compensation and classifier to enhance the relationship between the two stages. In this framework, we use the sparse coding (SC) to obtain the session-compensated feature by learning an overcomplete dictionary, and employ the softmax classifier and support vector machine (SVM) in classifying respectively. Moreover, we present a joint optimization of the dictionary and classifier parameters under a discriminative criterion for classifier with conditions for SC. In addition, the proposed methods are evaluated on the King-ASR-010, VoxCeleb and RSR2015 databases. Compared with typical session compensation techniques, such as linear discriminant analysis (LDA) and nonparametric discriminant analysis (NDA), our methods can be more robust to complex session variability. Moreover, compared with the typical classifiers in i-vector framework, i.e. the cosine distance scoring (CDS) and probabilistic linear discriminant analysis (PLDA), our methods can be more suitable for SID (multiclass task).     

Speaker-specific-text based speaker verification system using spectral and phase based features

In speaker recognition tasks, one of the reasons for reduced accuracy is due to closely resembling speakers in the acoustic space. In order to increase the discriminative power of the classifier, the system must be able to use only the unique features of a given speaker with respect to his/her acoustically resembling speaker. This paper proposes a technique to reduce the confusion errors, by finding speaker-specific phonemes and formulate a text using the subset of phonemes that are unique, for speaker verification task using i-vector based approach. In this paper spectral features such as linear prediction cepstral co-efficients (LPCC), perceptual linear prediction co-efficients (PLP) and phase feature such as modified group delay are experimented to analyse the importance of speaker-specific-text in speaker verification task. Experiments have been conducted on speaker verification task using speech data of 50 speakers collected in a laboratory environment. The experiments show that the equal error rate (EER) has been decreased significantly using i-vector approach with speaker-specific-text when compared to i-vector approach with random-text using different spectral and phase based features.     

基于深层置信网络的说话人信息提取方法

在基于全差异空间因子(i-Vector)的说话人确认系统中,需进一步从语音段的i-Vector表示中提取说话人相关的区分性信息,以提高系统性能。文中通过结合锚模型的思想,提出一种基于深层置信网络的建模方法。该方法通过对i-Vector中包含的复杂差异信息逐层进行分析、建模,以非线性变换的形式挖掘出其中的说话人相关信息。在NIST SRE 2008核心测试电话训练-电话测试数据库上,男声和女声的等错误率分别为4。96%和6。18%。进一步与基于线性判别分析的系统进行融合,能将等错误率降至4。74%和5。35%。     

基于Mellin变换的语音新特征与频率归正说话人自适应技术

为了减小由于说话人之间声道形状的差异而引起的非特定人语音识别系统性能的 下降,研究了两种方法,一种是基于最大似然估计的频率归正说话人自适应方法,另一种是基 于Mellin变换的语音新特征.在非特定人孤立词语音识别系统上的初步实验表明,这两种方 法都可以提高系统对不同说话人的鲁棒性,相比之下,基于Mellin变换的语音新特征具有更 好的性能,它不仅提高了系统对不同话者的识别性能,而且也使系统对不同话者的误识率的 离散程度大大减小.     

Mismatched feature detection with finer granularity for emotional speaker recognition

The shapes of speakers＇ vocal organs change under their different emotional states, which leads to the deviation of the emotional acoustic space of short-time features from the neutral acoustic space and thereby the degradation of the speaker recognition performance. Features deviating greatly from the neutral acoustic space are considered as mismatched features, and they negatively affect speaker recognition systems. Emotion variation produces different feature deformations for different phonemes, so it is reasonable to build a finer model to detect mismatched features under each phoneme. However, given the difficulty of phoneme recognition, three sorts of acoustic class recognition--phoneme classes, Gaussian mixture model （GMM） tokenizer, and probabilistic GMM tokenizer--are proposed to replace phoneme recognition. We propose feature pruning and feature regulation methods to process the mismatched features to improve speaker recognition performance. As for the feature regulation method, a strategy of maximizing the between-class distance and minimizing the within-class distance is adopted to train the transformation matrix to regulate the mismatched features. Experiments conducted on the Mandarin affective speech corpus （MASC） show that our feature pruning and feature regulation methods increase the identification rate （IR） by 3.64% and 6.77%, compared with the baseline GMM-UBM （universal background model） algorithm. Also, corresponding IR increases of 2.09% and 3.32% can be obtained with our methods when applied to the state-of-the-art algorithm i-vector.     

Improved i-vector extraction technique for speaker verification with short utterances

A major challenge in ASV is to improve performance with short speech segments for end-user convenience in real-world applications. In this paper, we present a detailed analysis of ASV systems to observe the duration variability effects on state-of-the-art i-vector and classical Gaussian mixture model-universal background model (GMM-UBM) based ASV systems. We observe an increase in uncertainty of model parameter estimation for i-vector based ASV with speech of shorter duration. In order to compensate the effect of duration variability in short utterances, we have proposed adaptation technique for Baum-Welch statistics estimation used to i-vector extraction. Information from pre-estimated background model parameters are used for adaptation method. The ASV performance with the proposed approach is considerably superior to the conventional i-vector based system. Furthermore, the fusion of proposed i-vector based system and GMM-UBM further improves the ASV performance, especially for short speech segments. Experiments conducted on two speech corpora, NIST SRE 2008 and 2010, have shown relative improvement in equal error rate (EER) in the range of 12–20%.     

高斯PLDA在说话人确认中的应用及其联合估计

近年来,基于总变化因子的说话人识别方法成为说话人识别领域的主流方法.其中,概率线性鉴别分析（Probabilistic linear discriminant analysis,PLDA）因其优异的性能而得到学者们的广泛关注.然而,在估计PLDA模型时,传统的因子分析方法只更新模型空间,因此,模型均值不能很好地与更新后的模型空间耦合.提出联合估计法对模型均值和模型空间同时估计,得到更为严格的期望最大化更新公式,在美国国家标准与技术局说话人识别评测2010扩展测试数据库以及2012核心测试数据库上,等错率得到一定提升.     

Speech / music classification using speech-specific features

This paper proposes the use of speech-specific features for speech / music classification. Features representing the excitation source, vocal tract system and syllabic rate of speech are explored. The normalized autocorrelation peak strength of zero frequency filtered signal, and peak-to-sidelobe ratio of the Hilbert envelope of linear prediction residual are the two source features. The log mel energy feature represents the vocal tract information. The modulation spectrum represents the slowly-varying temporal envelope corresponding to the speech syllabic rate. The novelty of the present work is in analyzing the behavior of these features for the discrimination of speech and music regions. These features are non-linearly mapped and combined to perform the classification task using a threshold based approach. Further, the performance of speech-specific features is evaluated using classifiers such as Gaussian mixture models, and support vector machines. It is observed that the performance of the speech-specific features is better compared to existing features. Additional improvement for speech / music classification is achieved when speech-specific features are combined with the existing ones, indicating different aspects of information exploited by the former.