Speaker Recognition Experiments Research Articles

PLDA inspired Siamese networks for speaker verification

The deep learning methodologies in state-of-the-art speaker recognition systems are predominantly limited to the extraction of recording level embeddings. This is usually followed by generative modeling of the embeddings to output the verification score. In this paper, we explore a fully neural approach where the neural model outputs the verification score directly, given the acoustic feature inputs. This model, termed as Siamese neural network (SiamNN), combines the embedding extraction and back-end modeling into a single processing pipeline. The back-end modeling is achieved using a neural approach to PLDA modeling, called neural probabilistic linear discriminant analysis (NPLDA). In the NPLDA model, the verification score is computed as a discriminative similarity function. The development of the single neural SiamNN model allows the joint optimization of all the modules using a verification cost. Several speaker recognition experiments are performed using SITW, VOiCES, and NIST SRE datasets where the proposed SiamNN model is shown to significantly improve over the state-of-art x-vector PLDA baseline system (relative improvements of up to 35% in the primary cost metric). We also provide a detailed analysis of the influence of hyper-parameters, choice of loss functions, and data sampling strategies for training the model. In particular, we highlight that the proposed soft detection cost function based optimization improves over other loss functions considered.

Text-independent speaker recognition based on adaptive course learning loss and deep residual network

Text-independent speaker recognition is widely used in identity recognition that has a wide spectrum of applications, such as criminal investigation, payment certification, and interest-based customer services. In order to improve the recognition ability of log filter bank feature vectors, a method of text-independent speaker recognition based on deep residual networks model was proposed in this paper. The deep residual network was composed of a residual network (ResNet) and a convolutional attention statistics pooling (CASP) layer. The CASP layer could aggregate frame-level features from the ResNet into an utterance-level features. Extracting speech features for each speaker using deep residual networks was a promising direction to explore, and a straightforward solution was to train the discriminative feature extraction network by using a margin-based loss function. However, a margin-based loss function often has certain limitations, such as the margins between different categories were set to be the same and fixed. Thus, we used an adaptive curriculum learning loss (ACLL) to address the problem and introduce two different margin-based losses for this problem, i.e., AM-Softmax and AAM-Softmax. The proposed method was applied to a large-scale VoxCeleb2 dataset for extensive text-independent speaker recognition experiments, and average equal error rate (EER) could achieve 1.76% on VoxCeleb1 test dataset, 1.91% on VoxCeleb1-E test dataset, and 3.24% on VoxCeleb1-H test dataset. Compared with related speaker recognition methods, EER was improved by 1.11% on VoxCeleb1 test dataset, 1.04% on VoxCeleb1-E test dataset, and 1.69% on VoxCeleb1-H test dataset.

Creation of a Nigerian Voice Corpus for Indigenous Speaker Recognition

One of the goals of Word Bank’s Identification for Development (ID4D) is the realization of robust digital identification systems as a means of sustainable development priority. ID4D’s most recent report shows about 1.1 billion of the world’s population are yet to be identified for development. Africa represents about half of that number while Nigeria represents about a quarter of Africa’s share. Biometrics is the state-of-the-art approach for identification using human behavioral and/or physiological digitally calibrated traits and one such trait is the voice. The backbone of biometric research is the database employed in the design of biometric systems. Although many voice databases are publicly available such as the THCHS-30 for Chinese and Microsoft Indian language Speech Corpus for Indians, none is currently publicly available or free for Nigerians. The creation of such an indigenous database (or corpus) can open doors to Nigerian automatic speaker recognition as well as for indigenous language, ethnicity, gender, age group and emotion classification amongst others. This work is a first step in the direction of creating a Nigerian Voice Corpus (NVC) to aid indigenous voice biometric research. A voice corpus of popular Nigerians was created by curation of audio samples of 14 women and 23 men from YouTube. The corpus contains 10 different samples of 5 seconds duration for each individual resulting in a total of 370 samples. The created corpus was used to carry out speaker recognition experiment by dividing the audio samples into 25ms non-overlapping frame durations. Silent frames were excluded using short-term spectral energy threshold for Voice Activity Detection (VAD). This was followed by extraction of Mel Frequency Cepstral Coefficient (MFCC) as descriptors to discriminate different speakers using Support Vector Machine (SVM) with median Gaussian function. An overall recognition accuracy of 93.24% was achieved demonstrating the feasibility and research potential in this direction.

Bone- and air-conduction speech combination method for speaker recognition

In this paper, first, we report speaker recognition performance using bone-conduction speech based on an i-vector-based speaker recognition system, which is the current state-of-the-art method. In addition, we propose three speaker recognition methods combining bone-conduction speech and air-conduction speech: a feature combination method, a speaker model combination method, and a similarity score combination method. To evaluate the proposed methods, we conducted speaker recognition experiments using a part of a large speech corpus constructed by the National Research Institute of Police Science, Japan. Experimental results show the bone-conduction speech performs almost the same as the air-conduction speech when the enrolment data and evaluation data are collected in the same session. In addition, all proposed methods improved the speaker recognition performance of air- and bone-conduction speech in the experiments. From these results, we conclude that fusing air- and bone-conduction speech improves the speaker recognition performance.

Effective use of combined excitation source and vocal-tract information for speaker recognition tasks

In automatic speaker recognition (SR) tasks the widely used score level combination scheme derives a general consensus from the independent opinions of individual evidences. Instead, we conjecture that collectively contributed decisions may be more effective. Based on this idea this work proposes an effective combination scheme, where the vocal-tract and excitation source information take decisions collectively, resulting higher improvements in SR accuracy. In the proposed scheme, independently made feature-specific models are padded for building resultant models. While testing, feature-specific test features are padded in similar fashion, and then used for comparison with resultant models. The main advantage of this proposed scheme is that it does not require any ground truth information for combined use of multiple evidences like in score level combination scheme. The potential of the proposed scheme is experimentally demonstrated by conducting different speaker recognition experiments in clean and noisy conditions, and also comparative studies with score level fusion scheme as reference. The TIMIT database is used for studies with clean case, and Indian Institute of Technology Guwahati Multi-Variability (IITG-MV) databases for noisy case. In clean case the proposed scheme provides relatively 1% of higher improvements in performance for GMM based speaker identification system and 8.5% for GMM–UBM based speaker verification system. In noisy case the corresponding parameters are 1% and 3%, respectively. The final evaluations on NIST-2003 database with GMM–UBM and i-vector based systems show relatively higher improvements in performance by 5.17% and 4.73%, respectively. The proposed scheme is observed to be statistically more significant than the commonly used score level fusion of multiple evidences.

Speaker recognition from whispered speech: A tutorial survey and an application of time-varying linear prediction

From the available biometric technologies, automatic speaker recognition is one of the most convenient and accessible ones due to abundance of mobile devices equipped with a microphone, allowing users to be authenticated across multiple environments and devices. Speaker recognition also finds use in forensics and surveillance. Due to the acoustic mismatch induced by varied environments and devices of the same speaker, leading to increased number of identification errors, much of the research focuses on compensating for such technology-induced variations, especially using machine learning at the statistical back-end. Another much less studied but at least as detrimental source of acoustic variation, however, arises from mismatched speaking styles induced by the speaker, leading to a substantial performance drop in recognition accuracy. This is a major problem especially in forensics where perpetrators may purposefully disguise their identity by varying their speaking style. We focus on one of the most commonly used ways of disguising one’s speaker identity, namely, whispering. We approach the problem of normal-whisper acoustic mismatch compensation from the viewpoint of robust feature extraction. Since whispered speech is intelligible, yet a low-intensity signal and therefore prone to extrinsic distortions, we take advantage of robust, long-term speech analysis methods that utilize slow articulatory movements in speech production. In specific, we address the problem using a novel method, frequency-domain linear prediction with time-varying linear prediction (FDLP-TVLP), which is an extension of the 2-dimensional autoregressive (2DAR) model that allows vocal tract filter parameters to be time-varying, rather than piecewise constant as in classic short-term speech analysis. Our speaker recognition experiments on the whisper subset of the CHAINS corpus indicate that when tested in normal-whisper mismatched conditions, the proposed FDLP-TVLP features improve speaker recognition performance by 7–10% over standard MFCC features in relative terms. We further observe that the proposed FDLP-TVLP features perform better than the FDLP and 2DAR methods for whispered speech.

Glottal flow 신호에서의 향상된 특징추출 및 다중 특징파라미터 결합을 통한 화자인식 성능 향상

In this paper, we utilize source mel-frequency cepstral coefficients (SMFCCs), skewness, and kurtosis extracted in glottal flow signals to improve speaker recognition performance. Generally, because the high band magnitude response of glottal flow signals is somewhat flat, the SMFCCs are extracted using the response below the predefined cutoff frequency. The extracted SMFCC, skewness, and kurtosis are concatenated with conventional feature parameters. Then, dimensional reduction by the principal component analysis (PCA) and the linear discriminat analysis (LDA) is followed to compare performances with conventional systems under equivalent conditions. The proposed recognition system outperformed the conventional system for large scale speaker recognition experiments. Especially, the performance improvement was more noticeable for small Gaussan mixtures.

Spectral shifting of speaker‐specific information for narrow band telephonic speaker recognition

Speech-based authentication system can perform remote authentication over telephone channels. However, telephone channels are restricted to a bandwidth of ∼0–4 kHz while studies on the distribution of speaker-specific information in the speech spectrum strongly suggests that useful speaker-specific information is present above 4 kHz. A method to shift a part of this speaker-specific information above 4 kHz into the telephone bandwidth in place of less speaker-specific information originally present below 4 kHz is proposed. Speaker recognition experiments conducted using the proposed method leads to ∼18.5% relative improvement on equal error rate when compared to a system using the conventional telephone band speech, as evaluated on the Intelligence Advanced Research Projects Activity (IARPA) Babel Program Tamil language collection.

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.

Composite Feature Selection Method based on Spoken Word and Speaker Recognition

The aim of this paper is to measure the recognition capability of composite features extracted from speech signal and compare the result with other individually considered features for both spoken word and speaker based recognitions. Standard features like formants (F1, F2, F3), Linear Predictive Coefficients (LPC) and Mel Frequency Cepstral Coefficients (MFCC) along with various combinations among them are considered for the task to arrive at the conclusion. Six different speakers and six different strings (words) are considered in the present study. The threshold is set through an iterative approach for both spoken word and speaker recognition experiments. The mixing of LPC and MFCC is found to be the most promising combination among all others. Another interesting conclusion that we can draw from the study that the composite feature approach gives accuracy very near to 100% in case of speaker recognition task as compared to spoken word recognition task.. General Terms Digital Signal Processing, Pattern Recognition, Artificial Neural Network, Automatic Speech and Speaker Recognition

Skew Gaussian mixture models for speaker recognition

Gaussian mixture models (GMMs) are widely used in speech and speaker recognition. This study explores the idea that a mixture of skew Gaussians might capture better feature vectors that tend to have skew empirical distributions. It begins with deriving an expectation maximisation (EM) algorithm to train a mixture of two-piece skew Gaussians that turns out to be not much more complicated than the usual EM algorithm used to train symmetric GMMs. Next, the algorithm is used to compare skew and symmetric GMMs in some simple speaker recognition experiments that use Mel frequency cepstral coefficients (MFCC) and line spectral frequencies (LSF) as the feature vectors. MFCC are one of the most popular feature vectors in speech and speaker recognition applications. LSF were chosen because they exhibit significantly more skewed distribution than MFCC and because they are widely used [together with the related immittance spectral frequencies (ISF)] in speech transmission standards. In the reported experiments, models with skew Gaussians performed better than models with symmetric Gaussians and skew GMMs with LSF compared favourably with both skew symmetric and symmetric GMMs that used MFCC.

Robust Feature Extraction Using Modulation Filtering of Autoregressive Models

Speaker and language recognition in noisy and degraded channel conditions continue to be a challenging problem mainly due to the mismatch between clean training and noisy test conditions. In the presence of noise, the most reliable portions of the signal are the high energy regions which can be used for robust feature extraction. In this paper, we propose a front end processing scheme based on autoregressive (AR) models that represent the high energy regions with good accuracy followed by a modulation filtering process. The AR model of the spectrogram is derived using two separable time and frequency AR transforms. The first AR model (temporal AR model) of the sub-band Hilbert envelopes is derived using frequency domain linear prediction (FDLP). This is followed by a spectral AR model applied on the FDLP envelopes. The output 2-D AR model represents a low-pass modulation filtered spectrogram of the speech signal. The band-pass modulation filtered spectrograms can further be derived by dividing two AR models with different model orders (cut-off frequencies). The modulation filtered spectrograms are converted to cepstral coefficients and are used for a speaker recognition task in noisy and reverberant conditions. Various speaker recognition experiments are performed with clean and noisy versions of the NIST-2010 speaker recognition evaluation (SRE) database using the state-of-the-art speaker recognition system. In these experiments, the proposed front-end analysis provides substantial improvements (relative improvements of up to 25%) compared to baseline techniques. Furthermore, we also illustrate the generalizability of the proposed methods using language identification (LID) experiments on highly degraded high-frequency (HF) radio channels and speech recognition experiments on noisy data.

Distant-talking speaker identification by generalized spectral subtraction-based dereverberation and its efficient computation

Abstract Previously, a dereverberation method based on generalized spectral subtraction (GSS) using multi-channel least mean-squares (MCLMS) has been proposed. The results of speech recognition experiments showed that this method achieved a significant improvement over conventional methods. In this paper, we apply this method to distant-talking (far-field) speaker recognition. However, for far-field speech, the GSS-based dereverberation method using clean speech models degrades the speaker recognition performance. This may be because GSS-based dereverberation causes some distortion between clean speech and dereverberant speech. In this paper, we address this problem by training speaker models using dereverberant speech obtained by suppressing reverberation from arbitrary artificial reverberant speech. Furthermore, we propose an efficient computational method for a combination of the likelihood of dereverberant speech using multiple compensation parameter sets. This addresses the problem of determining optimal compensation parameters for GSS. We report the results of a speaker recognition experiment performed on large-scale far-field speech with different reverberant environments to the training environments. The proposed GSS-based dereverberation method achieves a recognition rate of 92.2%, which compares well with conventional cepstral mean normalization with delay-and-sum beamforming using a clean speech model (49.0%) and a reverberant speech model (88.4%). We also compare the proposed method with another dereverberation technique, multi-step linear prediction-based spectral subtraction (MSLP-GSS). The proposed method achieves a better recognition rate than the 90.6% of MSLP-GSS. The use of multiple compensation parameters further improves the speech recognition performance, giving our approach a recognition rate of 93.6%. We implement this method in a real environment using the optimal compensation parameters estimated from an artificial environment. The results show a recognition rate of 87.8% compared with 72.5% for delay-and-sum beamforming using a reverberant speech model.

Exploiting Jump-Resonance Hysteresis in Silicon Auditory Front-Ends for Extracting Speaker Discriminative Formant Trajectories

Jump-resonance is a phenomenon observed in non-linear circuits where the amplitude of the output signal exhibits an abrupt jump when the frequency of the input signal is varied. For [Formula: see text] filters used in the design of analog auditory front-ends (AFEs), jump-resonance is generally considered to be undesirable and several techniques have been proposed in literature to avoid or alleviate this artifact. In this paper we explore the use of jump-resonance based hysteresis in [Formula: see text] band-pass filters for encoding speech formant trajectories. Using prototypes of silicon AFEs fabricated in a 0.5 μm CMOS process, we demonstrate the benefits of the proposed approach for extracting speaker discriminative features. These benefits are validated using speaker recognition experiments where consistent improvements in equal-error-rates (EERs) are achieved using the jump-resonance based features as compared to conventional features.

Noise-Robust Speaker Recognition Combining Missing Data Techniques and Universal Background Modeling

Although the field of automatic speaker recognition (ASR) has been the subject of extensive research over the past decades, the lack of robustness against background noise has remained a major challenge. This paper describes a noise-robust speaker recognition system that combines missing data (MD) recognition with the adaptation of speaker models using a universal background model (UBM). For MD recognition, the identification of reliable and unreliable feature components is required. For this purpose, the signal-to-noise ratio (SNR) based mask estimation performance of various state-of-the art noise estimation techniques and noise reduction schemes is compared. Speaker recognition experiments show that the usage of a UBM in combination with missing data recognition yields substantial improvements in recognition performance, especially in the presence of highly non-stationary background noise at low SNRs.

Computationally efficient frame-averaged FM feature extraction for speaker recognition

Recently, subband frame-averaged frequency modulation (FM) as a complementary feature to amplitude-based features for several speech based classification problems including speaker recognition has shown promise. One problem with using FM extraction in practical implementations is computational complexity. Proposed is a computationally efficient method to estimate the frame-averaged FM component in a novel manner, using zero crossing counts and the zero crossing counts of the differentiated signal. FM components, extracted from subband speech signals using the proposed method, form a feature vector. Speaker recognition experiments conducted on the NIST 2008 telephone database show that the proposed method successfully augments mel frequency cepstrum coefficients (MFCCs) to improve performance, obtaining 17% relative reductions in equal error rates when compared with an MFCC-based system.

Dependency of recognition rate on number of words for text-independent speaker recognition using vector quantization

In this research, we discuss speaker recognition using the Kohonen feature map. The map is constructed for each speaker, and it is trained by using log-power and fourteenth-order mel-frequency cepstral coefficients (MFCC) and their temporal difference. The quantization distortion is computed between the input speech and a specific vector on the feature map of each speaker. We conduct speaker recognition experiment based on VQ distortion. Utterances of prefectural name in Japan are used as speech data. We examine particularly the dependency of recognition rate on number of words used for recognition. According to our experiments of speaker identification, this system correctly recognizes 98.9% by using a single word for 40 male speakers, while it attains 100% by using more than three words. Moreover, we confirmed superiority of using VQ over HMM under the same experimental conditions.

Pseudo Pitch Synchronous Analysis of Speech With Applications to Speaker Recognition

The fine spectral structure related to pitch information is conveyed in Mel cepstral features, with variations in pitch causing variations in the features. For speaker recognition systems, this phenomenon, known as “pitch mismatch” between training and testing, can increase error rates. Likewise, pitch-related variability may potentially increase error rates in speech recognition systems for languages such as English in which pitch does not carry phonetic information. In addition, for both speech recognition and speaker recognition systems, the parsing of the raw speech signal into frames is traditionally performed using a constant frame size and a constant frame offset, without aligning the frames to the natural pitch cycles. As a result the power spectral estimation that is done as part of the Mel cepstral computation may include artifacts. Pitch synchronous methods have addressed this problem in the past, at the expense of adding some complexity by using a variable frame size and/or offset. This paper introduces Pseudo Pitch Synchronous (PPS) signal processing procedures that attempt to align each individual frame to its natural cycle and avoid truncation of pitch cycles while still using constant frame size and frame offset, in an effort to address the above problems. Text independent speaker recognition experiments performed on NIST speaker recognition tasks demonstrate a performance improvement when the scores produced by systems using PPS are fused with traditional speaker recognition scores. In addition, a better distribution of errors across trials may be obtained for similar error rates, and some insight regarding of role of the fundamental frequency in speaker recognition is revealed. Speech recognition experiments run on the Aurora-2 noisy digits task also show improved robustness and better accuracy for extremely low signal-to-noise ratio (SNR) data.

Fractal dimension of voice-signal waveforms

The fractal dimension is one important parameter that characterizes waveforms. In this paper, we derive a new method to calculate fractal dimension of digital voice-signal waveforms. We show that fractal dimension is an efficient tool for speaker recognition or speech recognition. It can be used to identify different speakers or distinguish speech. We apply our results to Chinese speaker recognition and numerical experiment shows that fractal dimension is an efficient parameter to characterize individual Chinese speakers. We have developed a semiautomatic voiceprint analysis system based on the theory of this paper and former researches.

Learning statistically efficient features for speaker recognition

We apply independent component analysis for extracting an optimal basis to the problem of finding efficient features for representing speech signals of a given speaker. The speech segments are assumed to be generated by a linear combination of the basis functions, thus the distribution of speech segments of a speaker is modeled by adapting the basis functions so that each source component is statistically independent. The learned basis functions are oriented and localized in both space and frequency, bearing a resemblance to Gabor wavelets. These features are speaker-dependent characteristics and to assess their efficiency we performed speaker recognition experiments and compared our results with the conventional Fourier basis. Our results show that the proposed method is more efficient than the conventional Fourier-based features, in that they can obtain a higher recognition rate and coding efficiency.