Forensic Automatic Speaker Recognition Research Articles

Validation of an ECAPA-TDNN system for Forensic Automatic Speaker Recognition under case work conditions

In this work, we tested different variants of a Forensic Automatic Speaker Recognition (FASR) system based on Emphasized Channel Attention, Propagation and Aggregation in Time Delay Neural Network (ECAPA-TDNN). To this scope, conditions reflecting those of a real forensic voice comparison case have been taken into consideration according to the forensic_eval_01 evaluation campaign settings. Using this recent neural model as an embedding extraction block, various normalization strategies at the level of embeddings and scores allowed us to observe the variations in system performance in terms of discriminating power, accuracy and precision metrics. Our findings suggest that the ECAPA-TDNN can be successfully used as a base component of a FASR system, managing to surpass the previous state of the art, at least in the context of the considered operating conditions.

An improved MMSE estimator based modified group delay spectrum for Forensic Automatic Speaker Recognition

This paper presents an improved speech enhancement algorithm called Minimum Mean Square Error (MMSE), based on the MODified Group Delay spectrum (MODGD), for Forensic Automatic Speaker Recognition (FASR) under noisy environments. This algorithm uses the MODGD instead of the amplitude spectrum, to compute the power spectrum of the noise-corrupt signal. In the proposed estimator, the MODGD retains most of the formants information. Therefore, it enhances the noisy speech signal with high quality even at extremely low Signal-to-Noise Ratio (SNR) levels. The evaluation of the improved algorithm in simulated FASR scenarios was performed by adding different noise levels, extracted from the NOISEX-92 database to the clean NIST2000-traces. The results obtained show that the proposed MMSE–MODGD estimator provides greater suppression of noise components in regions of low SNR than the MMSE estimator. In addition, there is a drastic reduction in Equal Proportion Probability (EPP) (the improvements are 1.84% for babble noise and 1.25% for factory and white noises), combining FASR techniques with the proposed MMSE–MODGD estimator than with the conventional estimator.

Classification of Pitch and Gender of Speakers for Forensic Speaker Recognition from Disguised Voices Using Novel Features Learned by Deep Convolutional Neural Networks

Voice disguise is a major concern in forensic automatic speaker recognition (FASR). Classifying the type of disguise is very important for speaker recognition. Pitch disguise is a very common type of disguise that criminals try to attempt. Among the different types of disguises, high pitch and low pitch voices show more distortion. The features that are robust for high pitch and low pitch voices are different. Moreover, the effect of disguise on male and female voices are also different. In this work, we classified high pitch and low pitch disguised voices for male and female voices using a novel set of features. We arranged Mel frequency cepstral coefficients (MFCC), ΔMFCC, and ΔΔMFCC features as three-dimensional features, and these are given as the RGB equivalent spectrogram inputs to pretrained AlexNet deep convolutional neural network (DCNN). We fused the AlexNet output features with corresponding MFCC correlation features. These fused features are the proposed novel features for disguise classification. Classification using neural network (NN) and support vector machine (SVM) classifiers are performed. Simulation results show that classification with SVM classifier using these novel features gives improved accuracy of 98.89% compared to 95.99% accuracy obtained by using DCNN output features using traditional spectrogram inputs.

The application of Gammatone frequency cepstral coefficients for forensic voice comparison under noisy conditions

ABSTRACT Compared with humans, who have more powerful auditory ability in discriminating and identifying speakers in noisy environments, traditional forensic automatic speaker recognizers do not perform well when dealing with noisy recordings. This paper proposes a GMM-UBM Forensic Automatic Speaker Recognition (FASR) System to reduce the effect of noise on performance. The system uses Gammatone Frequency Cepstral Coefficients (GFCC) based on an auditory periphery model and also incorporates a Principal Component Analysis (PCA) algorithm. The system was tested and validated using Mandarin voice databases compromised with different levels of white noise and office noise. The performance of the system was compared with a baseline system using Mel Frequency Cepstral Coefficients (MFCC) and also PCA under the same conditions. The results show that the performance of the combined GFCC system achieved a substantial improvement when compared with the baseline MFCC system under conditions of a high level of office noise.

Effect of covert recordings from vehicles on the performance of forensic automatic speaker recognition

Speech recordings made in vehicles with hidden recording devices (“bugs”) present a number of peculiarities compared to high-quality direct microphone recordings, and also ordinary mobile phone recordings. Inspired by a recent court case the present investigation was undertaken to quantitatively assess the effect of three types of covert recordings, including data transmission, on a widely-used forensic automatic speaker recognition (FASR) system (Batvox 3.1). Using simultaneous high-quality re-recording of speech samples from 50 speakers played back inside four automobiles and one 5-ton RV as a benchmark, equal-error rates (EERs) between zero and 2 % were found. The latter number was obtained for GSM-transmitted voice data. When the acoustic data were not transmitted but stored inside the covert recording device, or recorded and stored on a smartphone placed inside the vehicles, EERs between zero to 0.45 % were obtained. Generally, EERs are similar to those obtained using the same FASR system in studies with non-covert recordings (direct recordings, landline and mobile telephone). No effect of the type or model of the vehicles on EERs was observed.

Forensic comparison of ageing voices from automatic and auditory perspectives

Comparison of non-contemporary speech samples occurs frequently in forensic speaker-recognition cases. While the ageing-related changes in the voice have been well investigated, the effect of ‘vocal ageing’ on forensic speaker recognition has yet to be fully established. In this article, auditory and automatic experiments providing a deeper insight into the impact of ageing on forensic speaker recognition are presented. A listener test investigating the extent to which vocal ageing is detectable by lay listeners is first presented. A test set of 10 males and 10 females, with recordings spanning approximately 30 years per speaker, are taken from the Trinity College Dublin Speaker Ageing (TCDSA) database. Correct detection of ageing in two samples of the same speaker is found to increase from 64% at a 10-year age difference to 86% at a 30-year age difference. Ageing is significantly more detectable in female speakers than male speakers, and female listeners are significantly better at detecting ageing than male listeners. A link between ageing detectability and speaking fundamental frequency is also observed. A forensic automatic speaker recognition (FASR) experiment with ageing speakers is then presented. Given a test set of five male speakers from the TCDSA database, each with multiple recordings spanning 30–50 years, ageing is shown to progressively weaken the strength-of-evidence (likelihood ratios) of same-speaker comparisons. While there is inter-speaker variability in the extent of the ageing effect, instances of erroneous support for the different-speaker hypothesis are introduced for all speakers within a time-lapse of 10 years. The detrimental effect of ageing on the overall FASR system is also illustrated via Tippett plots.

Performance of likelihood ratios considering bounds on the probability of observing misleading evidence

In this article, we introduce a new tool, namely 'Limit Tippett Plots', to assess the performance of likelihood ratios in evidence evaluation including theoretical bounds on the probability of observing misleading evidence. To do that, we first review previous work about such bounds. Then we derive 'Limit Tippett Plots' that complements Tippett plots with information about the limits on the probability of observing misleading evidence, which are taken as a reference. Thus, a much richer way to measure performance of likelihood ratios is given. Finally, we present an experimental example in forensic automatic speaker recognition following the protocols of the Acoustics Laboratory of Guardia Civil, where it can be seen that 'Limit Tippett Plots' help to detect problems in the calculation of likelihood ratios. Language: en

Forensic Automatic Speaker Recognition Based on Likelihood Ratio Using Acoustic-phonetic Features Measured Automatically

Forensic speaker recognition is experiencing a remarkable paradigm shift in terms of the evaluation framework and presentation of voice evidence. This paper proposes a new method of forensic automatic speaker recognition using the likelihood ratio framework to quantify the strength of voice evidence. The proposed method uses a reference database to calculate the within- and between-speaker variability. Some acoustic-phonetic features are extracted automatically using the software VoiceSauce. The effectiveness of the approach was tested using two Mandarin databases: A mobile telephone database and a landline database. The experiment's results indicate that these acoustic-phonetic features do have some discriminating potential and are worth trying in discrimination. The automatic acoustic-phonetic features have acceptable discriminative performance and can provide more reliable results in evidence analysis when fused with other kind of voice features.

Forensic Automatic Speaker Recognition with Degraded and Enhanced Speech

Forensic Automatic Speaker Recognition with Degraded and Enhanced Speech

Value and interpretation of biometric evidence in forensic automatic speaker recognition.

Forensic speaker recognition is the process of determining if a specific individual (suspected speaker) is the source of a questioned voice recording (trace). Forensic automatic speaker recognition has proven an effective tool in the fight against crime, yet there is a constant need for more research due to the difficulties in adapting automatic methods of voice comparison to the forensic methodology that provides a coherent way of assessing and presenting recorded speech as scientific evidence. The ongoing paradigm shift in the forensic speaker recognition needs biometric methods for the calculation of the evidence value, its strength, and the evaluation of this strength under operating conditions of the casework. In such methods, the biometric evidence consists of the quantified degree of similarity between speaker-dependent features extracted from the trace and speaker-dependent features extracted from recorded speech of a suspect, represented by his/her model. This presentation aims at introducing deterministic and statistical automatic speaker recognition methods that provide several ways of quantifying and presenting recorded voice as biometric evidence, as well as the assessment of its strength (likelihood ratio) in the Bayesian interpretation framework, including scoring and direct methods, compatible with interpretations in other forensic disciplines.

High-performance session variability compensation in forensic automatic speaker recognition.

Recently, the main performance improvement in automatic speaker recognition technology has been due to session variability compensation techniques, mainly based on factor analysis (FA), which have reduced the equal error rate (EER) of state-of-the-art systems by a factor of 10 in less than five years (e.g., EER<2% for NIST SRE 2008 telephone speech). Moreover, such systems are able to compute millions of comparisons a thousand times faster than real time after speech features are extracted. However, some challenges remain, because if there is a mismatch between the conditions of the FA training database and the speech used for comparison, the effectiveness of the compensation significantly decreases. This problem is especially relevant in forensic voice comparison, where the availability of speech matching operational conditions is usually sparse. In this presentation we show the impact of this effect in realistic simulated case studies. We use the Baeza–Ahumada IV database, which contains speech acquired with the Spanish Guardia Civil facilities, used in their daily work. We also present algorithms to handle sparsity in the data used for training FA models. Finally, we outline future research plans in order to improve session variability compensation performance in forensically realistic conditions.

Accuracy assessment methods for likelihood ratio-based evidence evaluation

There is an increasing interest in scientific assessment methods for measuring the accuracy of the evidence evaluation process in any discipline in forensic science. The reasons for such interest are multiple, but the American Daubert rules have played an important role in motivating the forensic laboratories to find a proper way of expressing theaccuracyof their techniques [1]. Regarding theevidenceevaluationand interpretation process, and usually aimed at reporting figures of merit to a factfinder, the assessment frameworks proposed in the literature have been many, but some of them have become particularly popular. In this contribution,we reviewsome frequentlyusedmethods for the assessment of the evidence evaluation process, assuming that the weight of such evidence is expressed in the form of likelihood ratios. Threemethods used in the literature are described and compared: rates ofmisleading evidence (frequently found as false positive and negative rates); hypothesisdependent likelihood-ratio histograms and its cumulative version, Tippett plots; andmethodsbasedonstrictlyproper scoring rules,withemphasis in the information-theoretical Empirical Cross-Entropy (ECE). Moreover, we present a novel assessment method to be used jointly with Tippett plots, and based on theoretical bounds on the probability of observing misleading evidence [2]. This method, which has been coined ‘Limit Tippett plots’, is especially adequate to detect inaccuracies in the models used to represent the data, demonstrating its usefulness for any likelihood ratio-based evidence evaluation technique, and showing to incorporate rich and useful information to conventional Tippett plots. Results are contributed in two ways. First, some illustrating examples are generated using synthetic data, where typical errors in the evidence evaluation process are simulated. Second, some results are presented using forensic automatic speaker-recognition systems. On the one hand, both sets of results demonstrate the adequacy of the presented method, Limit Tippet plots, to detect inaccuracies in the models used for likelihood ratio computation. On the other hand, the novel proposed technique is compared to other assessment frameworks, and some discussion is presented about whether themethods are complementary in daily forensic work andwith regards to reporting to a factfinder.

Forensic automatic speaker recognition using Bayesian interpretation and statistical compensation for mismatched conditions

Awarding Institution: Ecole Polytechnique Fédérale de Lausanne (EPFL),
 Switzerland
 
 Date of Award: November 2005

Voice disguise and automatic speaker recognition

In this paper a newly developed Forensic Automatic Speaker Recognition System (FASRS) was introduced and the effect of 10 types of voice disguises that are common in forensic casework on the performance of this system was studied. In this study 10 types of disguised voices and normal voices from 20 male college students were used as test samples. Each disguised voice was compared with all normal voices in the database to make speaker identification and speaker verification. The result of speaker recognition is summarized and the influence of voice disguises on the FASRS is evaluated.

Forensic Automatic Speaker Recognition [Exploratory DSP

This article discusses the approaches, potential, and limitations of forensic automatic speaker recognition (FASR) which is an exploratory application of digital speech signal processing and pattern recognition for judicial purposes, particularly for law enforcement

Aural and automatic forensic speaker recognition in mismatched conditions

In this article, we compare aural and automatic speaker recognition in the context of forensic analyses, using a Bayesian framework for the interpretation of evidence.We use perceptual tests performed by non-experts and compare their performance with that of an automatic speaker recognition system. These experiments are performed with 90 phonetically untrained subjects. Several forensic cases were simulated, using the Polyphone IPSC-02 database, varying in linguistic content and technical conditions of recording. We estimate the strength of evidence for both humans and the baseline automatic system, calculating likelihood ratios using perceptual scores for humans and log-likelihood scores for the automatic system. A methodology analogous to the Bayesian interpretation in forensic automatic speaker recognition is applied to the perceptual scores given by humans in order to estimate the strength of evidence. The degradation of the accuracy of human recognition in mismatched recording conditions is contrasted with that of the automatic system under similar recording conditions. The conditions considered are fixed telephone, cellular telephone and noisy speech in forensically realistic conditions. The perceptual cues that the human subjects use to perceive differences in voices are studied, along with their importance in different recording conditions. We observe that while automatic speaker recognition shows higher accuracy in matched conditions of training and testing, its performance degrades significantly in mismatched conditions. Aural recognition accuracy is also observed to degrade from matched conditions to mismatched conditions and in mismatched conditions, the baseline automatic systems showed comparable or slightly degraded performance compared to the aural recognition systems. The baseline automatic system with adaptation to noisy conditions showed comparable or better performance than aural recognition. The higher level perceptual cues used by human listeners in order to recognise speakers are discussed. We also discuss the possibility of increasing the accuracy of automatic systems using the perceptual cues that remain robust to mismatched recording conditions.

On compensation of mismatched recording conditions in the Bayesian approach for forensic automatic speaker recognition

This paper deals with a procedure to compensate for mismatched recording conditions in forensic speaker recognition, using a statistical score normalization. Bayesian interpretation of the evidence in forensic automatic speaker recognition depends on three sets of recordings in order to perform forensic casework: reference ( R) and control ( C) recordings of the suspect, and a potential population database ( P), as well as a questioned recording ( QR) [1]. The requirement of similar recording conditions between suspect control database ( C) and the questioned recording ( QR) is often not satisfied in real forensic cases. The aim of this paper is to investigate a procedure of normalization of scores, which is based on an adaptation of the Test-normalization (T-norm) [2] technique used in the speaker verification domain, to compensate for the mismatch. Polyphone IPSC-02 database and ASPIC (an automatic speaker recognition system developed by EPFL and IPS-UNIL in Lausanne, Switzerland) were used in order to test the normalization procedure. Experimental results for three different recording condition scenarios are presented using Tippett plots and the effect of the compensation on the evaluation of the strength of the evidence is discussed.