Large-scale speaker identification

This paper presents a novel algorithm for fast nearest neighbor search. At the preprocessing stage, the proposed algorithm constructs a lower bound tree by agglomeratively clustering the sample points in the database. Calculation of the distance between the query and the sample points can be avoided if the lower bound of the distance is already larger than the minimum distance. The search process can thus be accelerated because the computational cost of the lower bound which can be calculated by using the internal node of the lower bound tree, is less than that of the distance. To reduce the number of the lower bounds actually calculated the winner-update search strategy is used for traversing the tree. Moreover, the query and the sample points can be transformed for further efficiency improvement. Our experiments show that the proposed algorithm can greatly speed up the nearest neighbor search process. When applying to the real database used in Nayar's object recognition system, the proposed algorithm is about one thousand times faster than the exhaustive search.

Nearest Neighbor search is one of the simplest and most intuitive ideas in data mining. Due to it's simplicity and diverse utility, Nearest Neighbor search is often found to be the workhorse of a variety of data mining, machine learning, and computer vision algorithms. For very high dimensional data, the naive linear search tends to be optimal. This is due to the so called curse of dimensionality in database search. To accelerate the search time, may researchers have turned to the power and efficiency of GPGPU computing; mainly for computing the data intensive distance metrics. Though promising speedups are achieved through this method, it fails to capitalize on many of the recent advances in algorithms for bounded error approximate nearest neighbor searches. Approximate nearest neighbor methods give satisfactory results, especially in very large and possibly redundant datasets, while boasting sub-linear search complexities. In this paper we present a c-approximate r-nearest neighbor search algorithm for CUDA using Locality Sensitive Hash with Nearest Neighbor search (LSH-NN). We implement this system in CUDA and test it on real world image SIFT vector data. Our tests show that we are able to achieve significant speedup over the serial version, with good approximations on scalability, while achieving the near-optimal search complexity of LSH-NN.

Recent work reported on fast speaker search over large speech data corpora has focused on using locality sensitive hashing (LSH) search with hashing functions approximating i-vector based cosine distances (CosDist) for model comparisons. Because of the superior performance of probabilistic linear discriminant analysis (PLDA) model reported on speaker identification (SID) in recent years, in this paper we focus on using PLDA for fast speaker search. It is challenging to approximate PLDA well with simple hashing functions, resulting in difficulty to combine it with LSH search. As an alternative, we adopt a clustering-based pruning strategy to speed up PLDA search. Our results show the strategy can significantly speed up search with minimal performance loss. Another focus of this work is on PLDA model adaptation to mismatched conditions under which the fast search runs. The technique we adopt to adapt the PLDA model is based on the LDA adaptation method reported in [1], primarily adapting the LDA transform. Our results show this adaptation improves PLDA performance significantly (over 25% relative) on data collected in different conditions. Our speed-up experiments running with adapted LDA show that gains from the adapted PLDA are retained after the speed-up.

This paper presents a fast Nearest Neighbor Search (NNS) algorithm for geometric data sets, based on a modified k-d tree using the cache technique. First, the tree-based search process of our algorithm starts from an appropriate (cached) leaf node, not from the root node. Initial search space is restricted within small limit around the leaf node and the recursive depth-first tentative search is excluded. Therefore, the length of the node traversal path can be shortened. Second, we developed several techniques of what and how information is cached and reused. The indexing sequence of data can be good information, as well as the data itself. Generally speaking, data are stored in consecutive order. So the indexing data can be similar to (i) the previous indexing data , (ii) the same indexing data in another data set, and (iii) even the NNS result of the previous iteration in the Iterative Closest Point algorithm. Furthermore, we introduce a new method to apply the proposed search algorithm to even randomly sequenced data. We evaluated the algorithm with three kinds of data set (LIDAR, Kinect, and randomly generated data), and the results show that the proposed algorithm is about five times faster than the conventional approximated NNS algorithm using k-d tree.

In this chapter, we present a framework for speaker identification using locality sensitive hashing (LSH) and supervectors. Instead of GMM evaluation that we considered in the previous chapter, we transform the problem of speaker identification into string comparison. The motivation behind doing so is to perform speaker identification with privacy while having minimal computational overhead. Similar to the speaker verification framework of Chapter 6, we convert the utterances into supervector features [CampbellSRS et al., 2006] that are invariant with the length of the utterance. By applying the LSH transformation to the supervectors, we reduce the problem of nearest-neighbor classification into string comparison. As LSH is not privacy-preserving, we apply the cryptographic hash function to the LSH keys. This allows us to check if two LSH keys match without knowing their contents.

We propose a new algorithm for fast approximate nearest neighbor search based on the properties of ordered vectors. Data vectors are classified based on the index and sign of their largest components, thereby partitioning the space in a number of cones centered in the origin. The query is itself classified, and the search starts from the selected cone and proceeds to neighboring ones. Overall, the proposed algorithm corresponds to locality sensitive hashing in the space of directions, with hashing based on the order of components. Thanks to the statistical features emerging through ordering, it deals very well with the challenging case of unstructured data, and is a valuable building block for more complex techniques dealing with structured data. Experiments on both simulated and real-world data prove the proposed algorithm to provide a state-of-the-art performance.

The Mel-Frequency Cepstral Coefficients (MFCC) feature can be cast-off in speaker recognition. The process of feature extraction of the speech signal using Mel-Frequency Cepstral Coefficients (MFCC) feature vectors will generate an acoustic speech signal. Locality Sensitive Hashing (LSH) is frequently used as a classifier for Big Data related problems. In this research, we proposed a new model based on MFCC and LSH to integrate into speaker recognition model. The main returns of our newly proposed model are to get robustness, effective and accurate results in comparison with MFCC+GMM, LPCC+GMM and MFCC+PNN models. This model also contributes to the literature of Big Data. In this model, first, we extract the MFCC features from the wave file then we applied LSH classifier on extracted feature to transform into hash-table. Finally, the hash-tables of train and test wave files are compared and obtained 92.66% speaker recognition accuracy. We compared the accuracy ratio of proposed model with other traditional models namely MFCC+GMM, MFCC+PNN, and LPCC+GMM. Experimental results show that proposed model is more accurate and robust than traditional models and good for speaker recognition.

Speech feature extraction is the most significant step in any Automatic speaker recognition system. In the last 60 years a lot of research has gone into parametric representation of these speech features. Several techniques are currently being used for Automatic Speaker Recognition. Yet Automatic Speaker Recognition still remains a confront mainly due to variations in speaker's vocal tract with time and health, varying environmental conditions, disparities in the behavior and quality of speech recorders etc. MFCC is a extensively used technique in Automatic speaker recognition. In this paper the performance of MFCC technique was evaluated in a quiet environment. A speaker database containing 30 male and 30 female speakers was created. Two separate experiments were conducted for the performance evaluation of MFCC technique when applied to K means clustering. In the first case the speech features were directly matched. In the second case a VQ codebook was created by clustering the training features of these 60 speakers. A distortion easure based on the minimum Euclidean distance was used for speaker recognition. The failure rate of speaker recognition in first ase was found to be was found to be 10% while in the second case as found to be 14%. Matlab-7.10.0 was used for this study.

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.

Given a collection of objects and an associated similarity measure, the all-pairs similarity search problem asks us to find all pairs of objects with similarity greater than a certain user-specified threshold. In order to reduce the number of candidates to search, locality-sensitive hashing (LSH) based indexing methods are very effective. However, most such methods only use LSH for the first phase of similarity search—that is, efficient indexing for candidate generation. In this article, we present BayesLSH , a principled Bayesian algorithm for the subsequent phase of similarity search—performing candidate pruning and similarity estimation using LSH. A simpler variant, BayesLSH-Lite , which calculates similarities exactly, is also presented. Our algorithms are able to quickly prune away a large majority of the false positive candidate pairs, leading to significant speedups over baseline approaches. For BayesLSH, we also provide probabilistic guarantees on the quality of the output, both in terms of accuracy and recall. Finally, the quality of BayesLSH’s output can be easily tuned and does not require any manual setting of the number of hashes to use for similarity estimation, unlike standard approaches. For two state-of-the-art candidate generation algorithms, AllPairs and LSH, BayesLSH enables significant speedups, typically in the range 2 × --20 × for a wide variety of datasets. We also extend the BayesLSH algorithm for kernel methods—in which the similarity between two data objects is defined by a kernel function. Since the embedding of data points in the transformed kernel space is unknown, algorithms such as AllPairs which rely on building inverted index structure for fast similarity search do not work with kernel functions. Exhaustive search across all possible pairs is also not an option since the dataset can be huge and computing the kernel values for each pair can be prohibitive. We propose K-BayesLSH an all-pairs similarity search problem for kernel functions. K-BayesLSH leverages a recently proposed idea— kernelized locality sensitive hashing (KLSH)—for hash bit computation and candidate generation, and uses the aforementioned BayesLSH idea for candidate pruning and similarity estimation. We ran a broad spectrum of experiments on a variety of datasets drawn from different domains and with distinct kernels and find a speedup of 2 × --7 × over vanilla KLSH.

Voice biometrics are being implemented globally in large scale applications such as remote banking, government e-services, transportation and building security access, autonomous vehicles, and healthcare. They have been integrated in numerous apps, often coupled with face biometrics and artificial intelligence methods. Voice biometrics products and solutions must meet three key re-quirements for the success in their deployment: they must be highly trustable regarding privacy protection; easy to use and always be available. This edited book presents the state of the art in voice biometrics research and technologies including implementation and deployment challenges in terms of interoperability, scalability and performance, and security. The team of editors and chapter authors combine a wealth of expertise from academia and the industry. Topics covered include the fundamentals of voice biometrics; design of countermeasures for replay attack; attacker's perspective for voice biometrics; voice biometrics; speaker de-identification; performance evaluation of voice biometrics solutions; standardization of voice biometrics technology; industry perspectives; joining forces of voice and facial biometrics; and future trends and challenges in voice biometrics. Providing comprehensive coverage of the field of voice biometrics, this authoritative volume will be of great interest to researchers, scientists, engineers, practitioners and advanced students involved in the fields of security, biometrics, forensic sciences, human computer interaction, speech processing, acoustics, multimedia, pattern recognition, and privacy-preserving, digital signal processing and speech technologies. It will also be of interest to researchers and professionals working in law and criminology.

We present SocialSense , a collaborative smartphone based speaker and mood identification and reporting system that uses a user’s voice to detect and log his/her speaking and mood episodes. SocialSense collaboratively works with other phones that are running the app present in the vicinity to periodically send/receive speaking and mood vectors to/from other users present in a social interaction setting, thus keeping track of the global speaking episodes of all users with their mood. In addition, it utilizes a novel event-adaptive dynamic classification scheme for speaker identification which updates the speaker classification model every time one or more users enter or leave the scenario, ensuring a most updated classifier based on user presence. Evaluation of using dynamic classifiers shows that SocialSense improves speaker identification accuracy by 30% compared to traditional static speaker identification systems, and a 10% to 43% performance boost under various noisy environments. SocialSense also improves the mood classification accuracy by 4% to 20% compared to the baseline approaches. Energy consumption experiments show that its device daily lifetime is between 10-14 hours.

Nearest neighbor (NN) search is a fundamental issue in many computer applications, such as multimedia search, computer vision and machine learning. While this problem is trivial in low-dimensional search spaces, it becomes much more difficult in higher dimension because of the phenomenon known as the curse of dimensionality, where the complexity grows exponentially with dimension and the data tends to show strong correlations between dimensions. In this paper, we introduce a new hashing method to efficiently cope with this challenge. The idea is to split the search space into many subspaces based on a number of jointly-independent and uniformly-distributed circular random variables (CRVs) computed from the data points. Our method has been tested on datasets of local SIFT and global GIST features and was compared to locality sensitive hashing (LSH), Spherical Hashing methods (HD and SHD) and the fast library for approximate nearest neighbor (FLANN) matcher by using linear search as a baseline. The experimental results show that our method outperforms all state-of-the-art methods for the GIST features. For SIFT features, the results indicate that our method significantly reduces the search query time while preserving the search quality and outperforms FLANN for datasets of size less than 200 K points.

Nearest neighbor search in high dimensional spaces is an interesting and important problem which is relevant for a wide variety of applications, including multimedia information retrieval, data mining, and pattern recognition. For such applications, the curse of high dimensionality tends to be a major obstacle in the development of efficient search methods. This paper addresses the problem of designing an efficient algorithm for high dimensional nearest neighbor search using a priority queue. The proposed algorithm is based on a simple linear search algorithm and eliminates unnecessary arithmetic operations from distance computations between multidimensional vectors. Moreover, we propose two techniques, a dimensional sorting method and a PCA-based method, to accelerate multidimensional search. Experimental results indicate that our scheme scales well even for a very large number of dimensions.

Convolutional Neural Networks (CNNs) and its variants have achieved impressive performance when used for different speech processing tasks like spoken language identification, speaker verification, speech emotion recognition, etc. Conventionally, CNNs for speech applications consider input features from fixed duration speech segments as input. In this work, we attempt to consider features from complete speech signal as input to CNN. We propose to use spatial pyramid pooling (SPP) layer in CNN architecture to remove the fixed length constraint and to consider features from varying length speech signals as input to CNN for an end to end training. Proposed architecture also results in varying size set of feature maps from convolution layer. Further, we propose novel CNN-based segment-level pyramid match kernel (CNN-SLPMK) as dynamic kernel between a pair of varying size set of feature maps for the classification framework using support vector machines (SVMs) based classifier. We demonstrate that our proposed approach achieves comparable results with state-of-the-art techniques for speech emotion recognition task.