Minimum Detection Cost Function Research Articles

Subband fusion of complex spectrogram for fake speech detection

The phase information was shown useful in fake speech detection. However, the most common reason why phase-based features are not widely used is phase wrapping. This makes the original phase hard to model directly. Therefore, it remains a challenge how to utilize the phase information effectively. To address this issue, this paper proposes a novel subband fusion of the complex spectrogram method for fake speech detection. The complex spectrogram is used as the input feature, containing both amplitude and phase spectrogram. In addition, subbands of the complex spectrogram are modeled separately. The idea is motivated by the fact that each frequency band has a different effect on the fake speech detection task. Finally, to make full use of the subbands, the subband results are fused. Experimental results on the ASVspoof 2019 LA dataset show that our proposed system achieves an equal error rate (EER) of 0.68% and a minimum tandem detection cost function (min t-DCF) of 0.0224.

Voice Deepfake Detection Using the Self-Supervised Pre-Training Model HuBERT

In recent years, voice deepfake technology has developed rapidly, but current detection methods have the problems of insufficient detection generalization and insufficient feature extraction for unknown attacks. This paper presents a forged speech detection method (HuRawNet2_modified) based on a self-supervised pre-trained model (HuBERT) to improve detection (and address the above problems). A combination of impulsive signal-dependent additive noise and additive white Gaussian noise was adopted for data boosting and augmentation, and the HuBERT model was fine-tuned on different language databases. On this basis, the size of the extracted feature maps was modified independently by the α-feature map scaling (α-FMS) method, with a modified end-to-end method using the RawNet2 model as the backbone structure. The results showed that the HuBERT model could extract features more comprehensively and accurately. The best evaluation indicators were an equal error rate (EER) of 2.89% and a minimum tandem detection cost function (min t-DCF) of 0.2182 on the database of the ASVspoof2021 LA challenge, which verified the effectiveness of the detection method proposed in this paper. Compared with the baseline systems in databases of the ASVspoof 2021 LA challenge and the FMFCC-A, the values of EER and min t-DCF decreased. The results also showed that the self-supervised pre-trained model with fine-tuning can extract acoustic features across languages. And the detection can be slightly improved when the languages of the pre-trained database, and the fine-tuned and tested database are the same.

Multi-task deep cross-attention networks for far-field speaker verification and keyword spotting

Personalized voice triggering is a key technology in voice assistants and serves as the first step for users to activate the voice assistant. Personalized voice triggering involves keyword spotting (KWS) and speaker verification (SV). Conventional approaches to this task include developing KWS and SV systems separately. This paper proposes a single system called the multi-task deep cross-attention network (MTCANet) that simultaneously performs KWS and SV, while effectively utilizing information relevant to both tasks. The proposed framework integrates a KWS sub-network and an SV sub-network to enhance performance in challenging conditions such as noisy environments, short-duration speech, and model generalization. At the core of MTCANet are three modules: a novel deep cross-attention (DCA) module to integrate KWS and SV tasks, a multi-layer stacked shared encoder (SE) to reduce the impact of noise on the recognition rate, and soft attention (SA) modules to allow the model to focus on pertinent information in the middle layer while preventing gradient vanishing. Our proposed model demonstrates outstanding performance in the well-off test set, improving by 0.2%, 0.023, and 2.28% over the well-known SV model emphasized channel attention, propagation, and aggregation in time delay neural network (ECAPA-TDNN) and the advanced KWS model Convmixer in terms of equal error rate (EER), minimum detection cost function (minDCF), and accuracy (Acc), respectively.

Replay Speech Detection Based on Dual-Input Hierarchical Fusion Network

Speech anti-spoofing is a crucial aspect of speaker recognition systems and has received a great deal of attention in recent years. Deep neural networks have achieved satisfactory results in datasets with similar training and testing data distributions, but their generalization ability is limited in datasets with different distributions. In this paper, we proposed a novel dual-input hierarchical fusion network (HFN) to improve the generalization ability of our model. The network had two inputs (the original speech signal and the time-reversed signal), which increased the volume and diversity of the training data. The hierarchical fusion model (HFM) enabled more thorough fusion of information from different input levels and improved model performance by fusing the two inputs after speech feature extraction. We finally evaluated the results using the ASVspoof 2021 PA (Physical Access) dataset, and the proposed system achieved an Equal Error Rate (EER) of 24.46% and a minimum tandem Detection Cost Function (min t-DCF) of 0.6708 in the test set. Compared with the four baseline systems in the ASVspoof 2021 competition, the proposed system min t-DCF values were decreased by 28.9%, 31.0%, 32.6%, and 32.9%, and the EERs were decreased by 35.7%, 38.1%, 45.4%, and 49.7%, respectively.

The Role of Long-Term Dependency in Synthetic Speech Detection

Although much progress has been made in synthetic speech detection, there lacks comprehensive analysis of the essential differences between spoofed and genuine speech. We here utilize supervised contrastive loss originated from contrastive learning as an analytical tool to characterize the class similarity structure of ASVspoof 2019 logical access (LA) dataset, which shows that an ideal back-end classifier for synthetic speech detection should have the ability to capture long-term dependencies. Recently, Transformer has been found to have an excellent ability in learning long-term dependencies of input data. We hence propose a back-end classifier based on Transformer Encoder for synthetic speech detection. Convolution blocks are added before the Transformer Encoder, which leverages inductive biases to improve the generalization ability. Compared to two-dimensional convolution, one-dimensional convolution makes better architectural assumptions about the input speech features, which helps with modeling long-term dependencies and decreases the risk of overfitting. The proposed Transformer combined with one-dimensional convolution has fewer parameters than most existing back-end classifiers, and achieves an equal error rate of 1.06% and a minimum tandem detection cost function metric of 0.0345 when evaluated on ASVspoof 2019 LA dataset, which is one of the best models reported in the literature.

A Speaker Verification Method Based on TDNN–LSTMP

In speaker recognition, a robust recognition method is essential. This paper proposes a speaker verification method that is based on the time-delay neural network (TDNN) and long short-term memory with recurrent project layer (LSTMP) model for the speaker modeling problem in speaker verification. In this work, we present the application of the fusion of TDNN and LSTMP to the i-vector speaker recognition system that is based on the Gaussian mixture model-universal background model. By using a model that can establish long-term dependencies to create a universal background model that contains a larger amount of speaker information, it is possible to extract more feature parameters, which are speaker dependent, from the speech signal. We conducted experiments with this method on four corpora: two in Chinese and two in English. The equal error rate, minimum detection cost function and detection error tradeoff curve are used as criteria for system performance evaluation. The experimental results show that the TDNN–LSTMP/i-vector speaker recognition method outperforms the baseline system on both Chinese and English corpora and has better robustness.

Identity Vector Extraction Using Shared Mixture of PLDA for Short‐Time Speaker Recognition

The state-of-the-art speaker recognition system degrades performance rapidly dealing with short-time utterances. It is known to all that identity vectors (i-vectors) extracted from short utterances have large uncertainties and standard Probabilistic linear discriminant analysis (PLDA) method can not exploit this uncertainty to reduce the effect of duration variation. In this work, we use Shared mixture of PLDA (SM-PLDA) to remodel the i-vectors utilizing their uncertainties. SM-PLDA is an improved generative model with a shared intrinsic factor, and this factor can be regarded as an identity vector containing speaker indentification information. This identity vector can be modeled by PLDA. Experimental results are evaluated by both equal error rate and minimum detection cost function. The results conducted on the National institute of standards and technology (NIST) Speaker recognition evaluation (SRE) 2010 extended tasks show that the proposed method has achieved significant improvements compared with ivector/ PLDA and some other advanced methods.

Unsupervised crosslingual adaptation of tokenisers for spoken language recognition

Phone tokenisers are used in spoken language recognition (SLR) to obtain elementary phonetic information. We present a study on the use of deep neural network tokenisers. Unsupervised crosslingual adaptation was performed to adapt the baseline tokeniser trained on English conversational telephone speech data to different languages. Two training and adaptation approaches, namely cross-entropy adaptation and state-level minimum Bayes risk adaptation, were tested in a bottleneck i-vector and a phonotactic SLR system. The SLR systems using the tokenisers adapted to different languages were combined using score fusion, giving 7–18% reduction in minimum detection cost function (minDCF) compared with the baseline configurations without adapted tokenisers. Analysis of results showed that the ensemble tokenisers gave diverse representation of phonemes, thus bringing complementary effects when SLR systems with different tokenisers were combined. SLR performance was also shown to be related to the quality of the adapted tokenisers.

A General Bayesian Model for Speaker Verification

This paper presents a general Bayesian model for speaker verification tasks. It is a generative probability model. Due to its simple analytical property, a computationally efficient expectation-maximization algorithm can be derived to obtain the model parameters. A closedform solution, which allows the scalable size of enrollment set, is given in a full Bayesian way for making speaker verification decisions. Factor analysis technique is employed to model the speaker-specific components, then the redundant information in this model will be dropped. Experimental results are evaluated by both equal error rate and minimum detection cost function. The proposed approach shows promising results on the National institute of standards and technology (NIST) Speaker recognition evaluation (SRE) 2010 extended and 2012 core tasks. Significant improvement is obtained when comparing with Gaussian probabilistic linear discriminant analysis, especially under phone-call conditions and mismatched train-test channel conditions. Contrast experimental results with other popular generative probability models are also presented in this paper.

Regularized All-Pole Models for Speaker Verification Under Noisy Environments

Regularization of linear prediction based mel-frequency cepstral coefficient (MFCC) extraction in speaker verification is considered. Commonly, MFCCs are extracted from the discrete Fourier transform (DFT) spectrum of speech frames. In this paper, DFT spectrum estimate is replaced with the recently proposed regularized linear prediction (RLP) method. Regularization of temporally weighted variants, weighted LP (WLP) and stabilized WLP (SWLP) which have earlier shown success in speech and speaker recognition, is also introduced. A novel type of double autocorrelation (DAC) lag windowing is also proposed to enhance robustness. Experiments on the NIST 2002 corpus indicate that regularized all-pole methods (RLP, RWLP and RSWLP) yield large improvement on recognition accuracy under additive factory and babble noise conditions in terms of both equal error rate (EER) and minimum detection cost function (MinDCF).

A segment selection technique for speaker verification

The performance of speaker verification systems degrades considerably when the test segments are utterances of very short duration. This might be either due to variations in score-matching arising from the unobserved speech sounds of short speech utterances or the fact that the shorter the utterance, the greater the effect of individual speech sounds on the average likelihood score. In other words, the effects of individual speech sounds have not been cancelled out by a large number of speech sounds in very short utterances. This paper presents a score-based segment selection technique for discarding portions of speech that result in poor discrimination ability in a speaker verification task. Theory is developed to detect the most significant and reliable speech segments based on the probability that the test segment comes from a fixed set of cohort models. This approach, suitable for any duration of test utterance, reduces the effect of acoustic regions of the speech that are not accurately modelled due to sparse training data, and makes a decision based only on the segments that provide the best-matched scores from the segment selection algorithm. The proposed segment selection technique provides reductions in relative error rate of 22% and 7% in terms of minimum Detection Cost Function (DCF) and Equal Error Rate (EER) compared with a baseline used the segment-based normalization, when evaluated on the short utterances of NIST 2002 Speaker Recognition Evaluation dataset.

Efficient on-line signature recognition based on multi-section vector quantization

This paper proposes a multi-section vector quantization approach for on-line signature recognition. We have used a database of 330 users which includes 25 skilled forgeries performed by 5 different impostors. This database is larger than those typically used in the literature. Nevertheless, we also provide results from the SVC database. Our proposed system obtains similar results as the state-of-the-art online signature recognition algorithm, Dynamic Time Warping, with a reduced computational requirement, around 47 times lower. In addition, our system improves the database storage requirements due to vector compression, and is more privacy-friendly because it is not possible to recover the original signature using the codebooks. Experimental results reveal that our proposed multi-section vector quantization achieves a 98% identification rate, minimum Detection Cost Function value equal to 2.29% for random forgeries and 7.75% for skilled forgeries.

Using MMSE to improve session variability estimation

In this paper, the Session Variability Subspace Projection (SVSP) method based on model compensation for speaker verification was improved using the Minimum Mean Square Error (MMSE) criterion. The issue of SVSP is that the speaker's session-independent supervector is approximated by the average of all his or her session-dependent GMM-supervectors when estimating SVSP matrix. However, the error between them does obviously exist. Our goal is to minimise it using MMSE criterion. Compared with the original SVSP, the proposed method could achieve an error rate reduction of 6.7% for EER and 5.3% for minimum detection cost function over the NIST SRE 2006 1C4W-dataset.

Biometric dispersion matcher

In this paper we propose a new classifier called a dispersion matcher. Our proposal is especially well adapted to those scenarios where a large number of classes and a small number of samples per class are available for training. This is the situation of biometric systems where just three to five measures per person are acquired during enrollment. This is just the opposite situation of other pattern recognition applications where a small number of classes and a large amount of training samples are available, such as handwritten digit recognition (10 classes) for ZIP code identification. The dispersion matcher trains a quadratic discriminant classifier to solve the dichotomy “Do these two feature vectors belong to the same person?”. In this way, we solve an important set of topics: (a) we can classify an open world problem and we do not need to train the model again if a new user is added, (b) we find a natural solution for feature selection, (c) experimental results with a priori threshold provides good results. We evaluate the proposed system with hand-geometry and face recognition problems (identification and verification). In hand geometry, we get a minimum detection cost function (DCF) for verification of 0.21% and a maximum identification rate of 99.1%, which compares favorably with other state-of-the-art methods. In face verification we achieve 5.59% DCF and 92.77% identification rate, which also compares favorably with the literature.

On-line signature recognition based on VQ-DTW

This paper studies some pattern recognition algorithms for on-line signature recognition: vector quantization (VQ), nearest neighbor (NN), dynamic time warping (DTW) and hidden Markov models (HMM). We have used a database of 330 users which includes 25 skilled forgeries performed by five different impostors. This database is larger than the typical ones found in the literature. Experimental results reveal that our first proposed combination of VQ and DTW (by means of score fusion) outperforms the other algorithms (DTW, HMM) and achieves a minimum detection cost function (DCF) value equal to 1.37% for random forgeries and 5.42% for skilled forgeries. In addition, we present another combined DTW–VQ scheme which enables improvement of privacy for remote authentication systems, avoiding the submission of the whole original dynamical signature information (using codewords, instead of feature vectors). This system achieves similar performance than DTW.

Application of the mutual information minimization to speaker recognition/identification improvement

In this paper we propose the inversion of nonlinear distortions in order to improve the recognition rates of a speaker recognizer system. We study the effect of saturations on the test signals, trying to take into account real situations where the training material has been recorded in a controlled situation, but the testing signals present some mismatch with the input signal level (saturations). The experimental results for speaker recognition shows that a combination of several strategies can improve the recognition rates with saturated test sentences from 80% to 89.39%, while the results with clean speech (without saturation) is 87.76% for one microphone, and for speaker identification can reduce the minimum detection cost function with saturated test sentences from 6.42% to 4.15%, while the results with clean speech (without saturation) is 5.74% for one microphone and 7.02% for the other one.