Speaker Recognition Using LIRA Neural Networks

This article contains information from our investigation in the field of voice recognition. For this purpose, we created a voice database that contains different phrases in two languages, English and Spanish, for men and women. As a classifier, the LIRA (Limited Receptive Area) grayscale neural classifier was selected. The LIRA grayscale neural classifier was developed for image recognition tasks and demonstrated good results. Therefore, we decided to develop a recognition system using this classifier for voice recognition. From a specific set of speakers, we can recognize the speaker’s voice. For this purpose, the system uses spectrograms of the voice signals as input to the system, extracts the characteristics and identifies the speaker. The results are described and analyzed in this article. The classifier can be used for speaker identification in security system or smart buildings for different types of intelligent devices.

Performance Evaluation of Acoustic-Spectrographic Voice Identification Method in Native and Non-Native Speech

The paper deals with acoustic-spectrographic voice identification method in terms of its performance in non-native language speech. Performance evaluation is conducted by comparing the result of the analysis of recordings containing native language speech with recordings that contain foreign language speech. Our research is based on Tajik and Russian speech of Tajik native speakers due to the character of the criminal situation with drug trafficking. We propose a pilot experiment that represents a primary attempt enter the field.

Speaker Identification by Atomic Decomposition of Learned Features Using Computational Auditory Scene Analysis Principals in Noisy Environments

Speaker recognition is performed in high Additive White Gaussian Noise (AWGN) environments using principals of Computational Auditory Scene Analysis (CASA). CASA methods often classify sounds from images in the time-frequency (T-F) plane using spectrograms or cochleargrams as the image. In this paper atomic decomposition implemented by matching pursuit performs a transform from time series speech signals to the T-F plane. The atomic decomposition creates a sparsely populated T-F vector in “weight space” where each populated T-F position contains an amplitude weight. The weight space vector along with the atomic dictionary represents a denoised, compressed version of the original signal. The arraignment or of the atomic indices in the T-F vector are used for classification. Unsupervised feature learning implemented by a sparse autoencoder learns a single dictionary of basis features from a collection of envelope samples from all speakers. The approach is demonstrated using pairs of speakers from the TIMIT data set. Pairs of speakers are selected randomly from a single district. Each speak has 10 sentences. Two are used for training and 8 for testing. Atomic index probabilities are created for each training sentence and also for each test sentence. Classification is performed by finding the lowest Euclidean distance between then probabilities from the training sentences and the test sentences. Training is done at a 30dB Signal-to-Noise Ratio (SNR). Testing is performed at SNR’s of 0 dB, 5 dB, 10 dB and 30dB. The algorithm has a baseline classification accuracy of ~93% averaged over 10 pairs of speakers from the TIMIT data set. The baseline accuracy is attributable to short sequences of training and test data as well as the overall simplicity of the classification algorithm. The accuracy is not affected by AWGN and produces ~93% accuracy at 0dB SNR.

An Intelligent Text Independent Speaker Identification Using VQ-GMM Model Based Multiple Classifier System

Speaker Identification (SI) is the task of establishing identity of an individual based on his/her voice characteristics. The SI task is typically achieved by two-stage signal processing: training and testing. The training process calculates speaker specific feature parameters from the speech and generates speaker models accordingly. In the testing phase, speech samples from unknown speakers are compared with the models and classified. Even though performance of speaker identification systems has improved due to recent advances in speech processing techniques, there is still need of improvement. In this paper, a Closed-Set Tex-Independent Speaker Identification System (CISI) based on a Multiple Classifier System (MCS) is proposed, using Mel Frequency Cepstrum Coefficient (MFCC) as feature extraction and suitable combination of vector quantization (VQ) and Gaussian Mixture Model (GMM) together with Expectation Maximization algorithm (EM) for speaker modeling. The use of Voice Activity Detector (VAD) with a hybrid approach based on Short Time Energy (STE) and Statistical Modeling of Background Noise in the pre-processing step of the feature extraction yields a better and more robust automatic speaker identification system. Also investigation of Linde-Buzo-Gray (LBG) clustering algorithm for initialization of GMM, for estimating the underlying parameters, in the EM step improved the convergence rate and systems performance. It also uses relative index as confidence measures in case of contradiction in identification process by GMM and VQ as well. Simulation results carried out on voxforge.org speech database using MATLAB highlight the efficacy of the proposed method compared to earlier work.

A Hybrid GMM/SVM System for Text Independent Speaker Identification

This paper proposes a novel approach that combines statistical models and support vector machines. A hybrid scheme which appropriately incorporates the advantages of both the generative and discriminant model paradigms is described and evaluated. Support vector machines (SVMs) are trained to divide the whole speakers' space into small subsets of speakers within a hierarchical tree structure. During testing a speech token is assigned to its corresponding group and evaluation using gaussian mixture models (GMMs) is then processed. Experimental results show that the proposed method can significantly improve the performance of text independent speaker identification task. We report improvements of up to 50% reduction in identification error rate compared to the baseline statistical model.

Speaker Identification by Joint Statistical Characterization in the Log Gabor Wavelet Domain

Real world Speaker Identification (SI) application differs from ideal or laboratory conditions causing perturbations that leads to a mismatch between the training and testing environment and degrade the performance drastically. Many strategies have been adopted to cope with acoustical degradation; wavelet based Bayesian marginal model is one of them. But Bayesian marginal models cannot model the inter-scale statistical dependencies of different wavelet scales. Simple nonlinear estimators for wavelet based denoising assume that the wavelet coefficients in different scales are independent in nature. However wavelet coefficients have significant inter-scale dependency. This paper enhances this inter-scale dependency property by a Circularly Symmetric Probability Density Function (CS-PDF) related to the family of Spherically Invariant Random Processes (SIRPs) in Log Gabor Wavelet (LGW) domain and corresponding joint shrinkage estimator is derived by Maximum a Posteriori (MAP) estimator. A framework is proposed based on these to denoise speech signal for automatic speaker identification problems. The robustness of the proposed framework is tested for Text Independent Speaker Identification application on 100 speakers of POLYCOST and 100 speakers of YOHO speech database in three different noise environments. Experimental results show that the proposed estimator yields a higher improvement in identification accuracy compared to other estimators on popular Gaussian Mixture Model (GMM) based speaker model and Mel-Frequency Cepstral Coefficient (MFCC) features.

Improved Weighted Matching for Speaker Recognition

Matching algorithms have significant importance in speaker recognition. Feature vectors of the unknown utterance are compared to feature vectors of the modeled speakers as a last step in speaker recognition. A similarity score is found for every model in the speaker database. Depending on the type of speaker recognition, these scores are used to determine the author of unknown speech samples. For speaker verification, similarity score is tested against a predefined threshold and either acceptance or rejection result is obtained. In the case of speaker identification, the result depends on whether the identification is open set or closed set. In closed set identification, the model that yields the best similarity score is accepted. In open set identification, the best score is tested against a threshold, so there is one more possible output satisfying the condition that the speaker is not one of the registered speakers in existing database. This paper focuses on closed set speaker identification using a modified version of a well known matching algorithm. The results of new matching algorithm indicated better performance on YOHO international speaker recognition database.

Evaluation of a Multi-Resolution Dyadic Wavelet Transform Method for usable Speech Detection

Many applications of speech communication and speaker identification suffer from the problem of co-channel speech. This paper deals with a multi-resolution dyadic wavelet transform method for usable segments of co-channel speech detection that could be processed by a speaker identification system. Evaluation of this method is performed on TIMIT database referring to the Target to Interferer Ratio measure. Co-channel speech is constructed by mixing all possible gender speakers. Results do not show much difference for different mixtures. For the overall mixtures 95.76% of usable speech is correctly detected with false alarms of 29.65%.

Improved Text-Independent Speaker Identification using Fused MFCC and IMFCC Feature Sets based on Gaussian Filter

A state of the art Speaker Identification (SI) system requires a robust feature extraction unit followed by a speaker modeling scheme for generalized representation of these features. Over the years, Mel-Frequency Cepstral Coefficients (MFCC) modeled on the human auditory system has been used as a standard acoustic feature set for speech related applications. On a recent contribution by authors, it has been shown that the Inverted Mel- Frequency Cepstral Coefficients (IMFCC) is useful feature set for SI, which contains complementary information present in high frequency region. This paper introduces the Gaussian shaped filter (GF) while calculating MFCC and IMFCC in place of typical triangular shaped bins. The objective is to introduce a higher amount of correlation between subband outputs. The performances of both MFCC & IMFCC improve with GF over conventional triangular filter (TF) based implementation, individually as well as in combination. With GMM as speaker modeling paradigm, the performances of proposed GF based MFCC and IMFCC in individual and fused mode have been verified in two standard databases YOHO, (Microphone Speech) and POLYCOST (Telephone Speech) each of which has more than 130 speakers.

Speaker Identification using Neural Networks

The speech signal conveys information about the identity of the speaker. The area of speaker identification is concerned with extracting the identity of the person speaking the utterance. As speech interaction with computers becomes more pervasive in activities such as the telephone, financial transactions and information retrieval from speech databases, the utility of automatically identifying a speaker is based solely on vocal characteristic. This paper emphasizes on text dependent speaker identification, which deals with detecting a particular speaker from a known population. The system prompts the user to provide speech utterance. System identifies the user by comparing the codebook of speech utterance with those of the stored in the database and lists, which contain the most likely speakers, could have given that speech utterance. The speech signal is recorded for N speakers further the features are extracted. Feature extraction is done by means of LPC coefficients, calculating AMDF, and DFT. The neural network is trained by applying these features as input parameters. The features are stored in templates for further comparison. The features for the speaker who has to be identified are extracted and compared with the stored templates using Back Propogation Algorithm. Here, the trained network corresponds to the output; the input is the extracted features of the speaker to be identified. The network does the weight adjustment and the best match is found to identify the speaker. The number of epochs required to get the target decides the network performance.

Speaker Identification Using Admissible Wavelet Packet Based Decomposition

Mel Frequency Cepstral Coefficient (MFCC) features are widely used as acoustic features for speech recognition as well as speaker recognition. In MFCC feature representation, the Mel frequency scale is used to get a high resolution in low frequency region, and a low resolution in high frequency region. This kind of processing is good for obtaining stable phonetic information, but not suitable for speaker features that are located in high frequency regions. The speaker individual information, which is non-uniformly distributed in the high frequencies, is equally important for speaker recognition. Based on this fact we proposed an admissible wavelet packet based filter structure for speaker identification. Multiresolution capabilities of wavelet packet transform are used to derive the new features. The proposed scheme differs from previous wavelet based works, mainly in designing the filter structure. Unlike others, the proposed filter structure does not follow Mel scale. The closed-set speaker identification experiments performed on the TIMIT database shows improved identification performance compared to other commonly used Mel scale based filter structures using wavelets.

Improved Closed Set Text-Independent Speaker Identification by Combining MFCC with Evidence from Flipped Filter Banks

A state of the art Speaker Identification (SI) system requires a robust feature extraction unit followed by a speaker modeling scheme for generalized representation of these features. Over the years, Mel-Frequency Cepstral Coefficients (MFCC) modeled on the human auditory system has been used as a standard acoustic feature set for SI applications. However, due to the structure of its filter bank, it captures vocal tract characteristics more effectively in the lower frequency regions. This paper proposes a new set of features using a complementary filter bank structure which improves distinguishability of speaker specific cues present in the higher frequency zone. Unlike high level features that are difficult to extract, the proposed feature set involves little computational burden during the extraction process. When combined with MFCC via a parallel implementation of speaker models, the proposed feature set outperforms baseline MFCC significantly. This proposition is validated by experiments conducted on two different kinds of public databases namely YOHO (microphone speech) and POLYCOST (telephone speech) with Gaussian Mixture Models (GMM) as a Classifier for various model orders.

Text-independent Speaker Identification Based on MAP Channel Compensation and Pitch-dependent Features

One major source of performance decline in speaker recognition system is channel mismatch between training and testing. This paper focuses on improving channel robustness of speaker recognition system in two aspects of channel compensation technique and channel robust features. The system is text-independent speaker identification system based on two-stage recognition. In the aspect of channel compensation technique, this paper applies MAP (Maximum A Posterior Probability) channel compensation technique, which was used in speech recognition, to speaker recognition system. In the aspect of channel robust features, this paper introduces pitch-dependent features and pitch-dependent speaker model for the second stage recognition. Based on the first stage recognition to testing speech using GMM (Gaussian Mixture Model), the system uses GMM scores to decide if it needs to be recognized again. If it needs to, the system selects a few speakers from all of the speakers who participate in the first stage recognition for the second stage recognition. For each selected speaker, the system obtains 3 pitch-dependent results from his pitch-dependent speaker model, and then uses ANN (Artificial Neural Network) to unite the 3 pitch-dependent results and 1 GMM score for getting a fused result. The system makes the second stage recognition based on these fused results. The experiments show that the correct rate of two-stage recognition system based on MAP channel compensation technique and pitch-dependent features is 41.7% better than the baseline system for closed-set test.