Effect of Child Voices on Forensic Voice Comparison using Deep Speaker Embeddings

In forensic voice comparison, deep learning has become widely popular recently. It is mainly used to learn speaker representations, called embeddings or embedding vectors. Speaker embeddings are often trained using corpora mostly containing widely spoken languages. Thus, language dependency is an important factor in automatic forensic voice comparison, especially when the target language is linguistically very different from that the model is trained on. In the case of a low-resource language, developing a corpus for forensic purposes containing enough speakers to train deep learning models is costly. This study aims to investigate whether a model pre-trained on multilingual (mostly English) corpus can be used on a target low-resource language (here, Hungarian), not represented by the model. Often multiple samples are not available from the offender (unknown speaker). Samples are therefore compared pairwise with and without speaker enrollment for suspect (known) speakers. Two corpora are used that were developed especially for forensic purposes and a third that is meant for traditional speaker verification. Speaker embedding vectors are extracted by the x-vector and ECAPA-TDNN techniques. Speaker verification was evaluated in the likelihood-ratio framework. A comparison is made between the language combinations (modeling, LR calibration, and evaluation). The results were evaluated by Cllrmin and EER metrics. It was found that the model pre-trained on a different language but on a corpus with a significant number of speakers can be used on samples with language mismatch. Sample duration and speaking style also seem to affect the performance.

The mismatch between the training conditions and the test conditions severely degrades the performance of speaker verification. Aiming at solving this problem, this paper presents a method which uses a combination of improved nonnegative matrix factorization (IMNMF) and mel frequency cepstral coefficients (MFCC) with feature warping for improving identity-vector (i-vector) speaker verification performance in noisy environment. Unlike the traditional nonnegative matrix factorization (NMF), IMNMF uses extra free basis vectors to capture the features which are not included in training data, and linear constraints on dictionary atoms. Feature warping is used to remove channel noises. Therefore, the proposed method can reduce distortion of reconstructed speech while enhancing the recovered speech quality. The performance of i-vector speaker verification is evaluated by using the short utterance database and the NOISEX-92 database. The experiment results indicate that the score level fusion of feature-warped IMNMF-MFCC and feature-warped MFCC is superior to the baseline system at the equal error rate under the majority of signal-to-noise ratios (SNRs).

In recent years, speaker verification technologies have received an extensive amount of attention. Designing and developing machines that could communicate with humans are believed to be one of the primary motivations behind such developments. Speaker verification technologies are applied to numerous fields such as security, Biometrics, and forensics. In this paper, the authors study the effects of different languages on the performance of the automatic speaker verification (ASV) system. The MirasVoice speech corpus (MVSC), a bilingual English and Farsi speech corpus, is used in this study. This study collects results from both an I-vector based ASV system and a GMM-UBM based ASV system. The experimental results show that a mismatch between the enrolled data used for training and verification data can lead to a significant decrease in the overall system efficiency. This study shows that it is best to use an i-vector based framework with data from the English language used in the enrollment phase to improve the robustness of the ASV systems. The achieved results in this study indicate that this can narrow the degradation gap caused by the language mismatch.

In this paper, we present a Mixture Linear Prediction based approach for robust Gammatone Cepstral Coefficients extraction (MLPGCCs). The proposed method provides performance improvement of Automatic Speaker Verification (ASV) using i-vector and Gaussian Probabilistic Linear Discriminant Analysis GPLDA modeling under transmission channel noise. The performance of the extracted MLPGCCs was evaluated using the NIST 2008 database where a single channel microphone recorded conversational speech. The system is analyzed in the presence of different channel transmission noises such as Additive White Gaussian (AWGN) and Rayleigh fading at various Signals to Noise Ratio (SNR) levels. The evaluation results show that the MLPGCCs features are a promising way for the ASV task. Indeed, the speaker verification performance using the MLPGCCs proposed features is significantly improved compared to the conventional Gammatone Frequency Cepstral Coefficients (GFCCs) and Mel Frequency Cepstral Coefficients (MFCCs) features. For speech signals corrupted with AWGN noise at SNRs ranging from (-5 dB to 15 dB), we obtain a significant reduction of the Equal Error Rate (EER) ranging from 9.41% to 6.65% and 3.72% to 1.50%, compared with conventional MFCCs and GFCCs features respectively. In addition, when the test speech signals are corrupted with Rayleigh fading channel we achieve an EER reduction ranging from 23.63% to 7.8% and from 10.88% to 6.8% compared with conventional MFCCs and GFCCs, respectively. We also found that the combination of GFCCs and MLPGCCs gives the highest performance of speaker verification system. The best performance combination achieved is around EER from 0.43% to 0.59% and 1.92% to 3.88%.

Over recent years, i-vector-based framework has been proven to provide state-of-the-art performance in speaker verification. Each utterance is projected onto a total factor space and is represented by a low-dimensional feature vector. Channel compensation techniques are carried out in this low-dimensional feature space. Most of the compensation techniques take the sets of extracted i-vectors as input. By constructing between-class covariance and within-class covariance, we attempt to minimize the between-class variance mainly caused by channel effect and to maximize the variance between speakers. In the real-world application, enrollment and test data from each user (or speaker) are always scarce. Although it is widely thought that session variability is mostly caused by channel effects, phonetic variability, as a factor that causes session variability, is still a matter to be considered. We propose in this paper a new i-vector extraction algorithm from the total factor matrix which we term component reduction analysis (CRA). This new algorithm contributes to better modelling of session variability in the total factor space. We reported results on the male English trials of the core condition of the NIST 2008 Speaker Recognition Evaluation (SREs) dataset. As measured both by equal error rate and the minimum values of the NIST detection cost function, 10–15 % relative improvement is achieved compared to the baseline of traditional i-vector-based system.

Recently, x-vector [1] has been a successful and popular approach for speaker verification, which employs a time delay neural network (TDNN) and statistics pooling to extract speaker characterizing embedding from variable-length utterances. Improvement upon the x-vector has been an active research area, and enormous neural networks have been elaborately designed based on the x-vector, e.g., extended TDNN (E-TDNN) [2], factorized TDNN (F-TDNN) [3], and densely connected TDNN (D-TDNN) [4]. In this work, we try to identify the optimal architectures from a TDNN based search space employing neural architecture search (NAS), named SpeechNAS. Leveraging the recent advances in the speaker recognition, such as high-order statistics pooling, multi-branch mechanism, D-TDNN and angular additive margin softmax (AAM) loss with a minimum hyper-spherical energy (MHE), SpeechNAS automatically discovers five network architectures, from SpeechNAS-1 to SpeechNAS-5, of various numbers of parameters and GFLOPs on the large-scale text-independent speaker recognition dataset VoxCelebl. Our derived best neural network achieves an equal error rate (EER) of 1.02% on the standard test set of VoxCelebl, which surpasses previous TDNN based state-of-the-art approaches by a large margin.

This is the author’s version of a work that was submitted/accepted for pub-lication in the following source:

In the field of text-independent speaker recognition, dynamic models that adapt along the time axis have been proposed to consider the phoneme-varying characteristics of speech. However, a detailed analysis of how dynamic models work depending on phonemes is insufficient. In this paper, we propose temporal dynamic CNN (TDY-CNN) that considers temporal variation of phonemes by applying kernels optimally adapting to each time bin. These kernels adapt to time bins by applying weighted sum of trained basis kernels. Then, an analysis of how adaptive kernels work on different phonemes in various layers is carried out. TDY-ResNet-38(×0.5) using six basis kernels improved an equal error rate (EER), the speaker verification performance, by 17.3% compared to the baseline model ResNet-38(×0.5). In addition, we showed that adaptive kernels depend on phoneme groups and are more phoneme-specific at early layers. The temporal dynamic model adapts itself to phonemes without explicitly given phoneme information during training, and results show the necessity to consider phoneme variation within utterances for more accurate and robust text-independent speaker verification.

Phonetic information is one of the most essential components of a speech signal, playing an important role for many speech processing tasks. However, it is difficult to integrate phonetic information into speaker verification systems since it occurs primarily at the frame level while speaker characteristics typically reside at the segment level. In deep neural network-based speaker verification, existing methods only apply phonetic information to the frame-wise trained speaker embeddings. To improve this weakness, this paper proposes phonetic adaptation and hybrid multi-task learning and further combines these into c-vector and simplified c-vector architectures. Experiments on National Institute of Standards and Technology (NIST) speaker recognition evaluation (SRE) 2010 show that the four proposed speaker embeddings achieve better performance than the baseline. The c-vector system performs the best, providing over 30% and 15% relative improvements in equal error rate (EER) for the core-extended and 10 s–10 s conditions, respectively. On the NIST SRE 2016, 2018, and VoxCeleb datasets, the proposed c-vector approach improves the performance even when there is a language mismatch within the training sets or between the training and evaluation sets. Extensive experimental results demonstrate the effectiveness and robustness of the proposed methods.

This work is dedicated to introducing, executing, and assessing a three-stage speaker verification framework to enhance the degraded speaker verification performance in emotional talking environments. Our framework is comprised of three cascaded stages: gender identification stage followed by an emotion identification stage followed by a speaker verification stage. The proposed framework has been assessed on two distinct and independent emotional speech datasets: our collected dataset and Emotional Prosody Speech and Transcripts dataset. Our results demonstrate that speaker verification based on both gender cues and emotion cues is superior to each of speaker verification based on gender cues only, emotion cues only, and neither gender cues nor emotion cues. The achieved average speaker verification performance based on the suggested methodology is very similar to that attained in subjective assessment by human listeners.

RESUMEN Currently, many speaker recognition applications must handle speech corrupted by environmental additiv e noise without having a priori knowledge about the charact eristics of noise. Some previous works in speaker r ecognition have used Missing Feature (MF) approach to compensate for noise. In most of those applications the spe ctral reliability decision step is done using the Signal to Noise Ratio (SNR) criterion. This has the goal o f enhancing signal power rather than noise power, which could b e dangerous in speaker recognition tasks, because u seful speaker information could be removed. This work proposes a new mask estimation method based on Speaker Discriminative Information (SDI) for determining spectral reliabil ity in speaker recognition applications based on th e MF approach. The proposal was evaluated through speaker verification experiments in speech corrupted by additive no ise. Experiments demonstrated that this new criterion ha s a promising performance in speaker verification t asks.

A method is described for estimating telephone handset nonlinearity by matching the spectral magnitude of the distorted signal to the output of a nonlinear channel model, driven by an undistorted reference. This "magnitude only" representation allows the model to directly match unwanted speech formants that arise over nonlinear channels and that are a potential source of degradation in speaker and speech recognition algorithms. As such, the method is particularly suited to algorithms that use only spectral magnitude information. The distortion model consists of a memoryless nonlinearity sandwiched between two finite-length linear filters. Nonlinearities considered include arbitrary finite-order polynomials and parametric sigmoidal functionals derived from a carbon-button handset model. Minimization of a mean-squared spectral magnitude distance with respect to model parameters relies on iterative estimation via a gradient descent technique. Initial work has demonstrated the importance of addressing handset nonlinearity, in addition to linear distortion, in speaker recognition over telephone channels. A nonlinear handset "mapping," applied to training or testing data to reduce mismatch between different types of handset microphone outputs, improves speaker verification performance relative to linear compensation only. Finally, a method is proposed to merge the mapper strategy with a method of likelihood score normalization (hnorm) for further mismatch reduction and speaker verification performance improvement.

The purpose of this paper is to preliminarily present the status of the paradigm shift or new paradigm in forensic voice comparison in Peru. First, the ambiguous terminology (such as forensic acoustics, voice identification or forensic phonoaudiology as forensic voice comparison) used to avoid its inappropriate use in academic and judicial situations is reviewed. Second, forensic phonetics and its areas of practice are presented in general terms. Third, the methodological approaches in voice comparison and the way of expressing the results are presented and it is explained why the forensic expert in voice comparison should opt for the analysis and results through the likelihood ratio. Fourth, the basis of the paradigm shift and its influence on forensic voice comparison is discussed in depth. Fifth, the current state of research in forensic voice comparison in Peru is presented. As a result of this research, it can be argued that the current studies demonstrate a transition towards a paradigm shift, especially through the use of automated programs.

This paper presents a comparison between several features normalization methods, and a comparison between different types of Gaussian Mixture Model (GMM) based supervectors normalizations for robust Speaker Verification. We implemented the methods of normalizations as a part of speaker verification system using Support Vector Machine (SVM) classifier and GMM-based supervectors. When implementing the speaker recognition system, we used Mel Frequency Cepstral Coefficients (MFCC) feature extraction. A valid question is which features normalization to use, if any. We examine the most common methods of feature normalizations, such as: Feature Warping mapping, and Cepstral Mean Subtraction (CMS) normalization with and without variance normalization. These methods were compared to features without normalization at all, and to a basic [-1, 1] normalization. In addition, we applied few types of normalizations to the GMM-mean supervectors, in order to improve the performance of the SVM classifier. All comparisons of the speaker verification system had been done in terms of DET curve, EER (Equal Error Rate) and Min. DCF. The best results we achieved were on combined supervector normalizations of Universal Background Model (UBM) Standard Deviation (STD) and [-1, 1] normalization. The type of the MFCC normalization has no big influence on the verification performance. The best results were: EER about 5.0% and MIN. DCF of 0.02.

Voice biometrics (e.g., Speaker Verification) is a critical type of biometrics based on human voice characteristics and is known for security and user-friendliness. It has been widely applied in worldwide applications, such as voice assistants and online banking. However, a concern is raised rapidly about the demographic fairness that different subgroups may have different speaker verification performance due to the inherent voice characteristics. And little work done investigates this concern. A diverse group of 300 speakers by race and gender is recruited for exploration. After running some speaker verification evaluations, three conclusions were reached. Firstly, the Latinx are performed the worst among the four major races in the US (White, Black, Latinx, and Asian) in speaker verification. Secondly, that gender shows little difference in performance between men and women. Thirdly, that high entropy voices performed better than low entropy voices in speaker verification performance.