Search

This report summarizes our effort towards building a robust open-set speaker recognition system. It reviews the various techniques we have used for acoustic feature extraction, speaker modeling, scoring and score normalization, and presents experiment results. We have worked on all the modules of speaker recognition systems. At the front end, we have studied a variety of acoustic features and pre-/post-processing techniques, and have come up with a PPMD feature that combines the benefits of multitaper MFCC, DSCC, pre-emphasis, and short-time feature Gaussianization. At the speaker modeling and scoring stages, we have investigated GMM speaker modeling, SVM speaker modeling, and joint factor analysis (JFA). We have demonstrated that compared to GMM modeling, SVM modeling and scoring are not only better and also faster. We have also shown that T-norm of the scores improves speaker identification performance on the ROSSI database., The original document contains color images.

Technological advances have had profound effects on the conduct of military operations in both peacetime and in war. One advance that has had a great impact outside the military by reducing human intervention is Voice Recognition (VR) technology. This thesis will examine the implementation of a Voice Recognition System as a shipdriving device and as a means of decreasing the occurrence of mishaps while reducing the level of fatigue of watchstanders on the bridge. Chapter I will discuss the need for the United States Navy to investigate the implementation of a Voice Recognition System to help reduce the probability of mishaps occurring. Chapter II will explain voice recognition technology, how it works, and how the proposed system can be fielded aboard U.S. Navy ships. Chapter III will examine the opinions (on the implementation of a Voice Recognition System) of officers charged with the safe navigation of naval ships. Chapter IV will review the concerns of officers, and will justify the implementation by answering these concerns. The conclusion will iterate the advances in voice recognition, and why a Voice Recognition system should be implemented on the bridges of U.S. Navy ships., http://archive.org/details/voicerecognition109451047, Lieutenant, United States Navy, Approved for public release; distribution is unlimited.

In February 2008, the Deputy Secretary of Defense signed a DoD Directive that established the Secretary of the Army as the DoD Executive Agent for DoD biometrics. The directive also indicated the importance of biometrics as a fully integrated enabling technology intended to support military operations. Even before that directive was signed, biometrics was being used extensively in a range of military operations. Despite its success, there has been little investigation of the potential use of biometrics in future operations. This report consists of two parts that summarize the conditions under which biometric collection may occur in future Army deployments. Part I describes a range of biometric modalities and discusses technical factors associated with their use in various environmental contexts. Part II describes social and anthropological considerations that lead to effective biometric collection., The original document contains color images.

The use of and dependence on, Information technology (IT) has grown tremendously in the last two decades. Still, some believe the United States is only in the infancy of this growth. This explosive growth has opened the door to capabilities that were only dreamed of in the past. As easy as it is to see how advantageous this technology is, it also is clear that with its advantages come distinct responsibilities and new problems that must be addressed. For instance, the minute one begins using information processing systems, the world of information assurance (IA) becomes far more complex. As a result, the push for better IA is necessary. To reach this increased level of IA, a further dependence on technology has developed. As an example, the field of biometrics has matured and has become an enabler to the U.S. Department of Defense IA model., The original document contains color images.

In the Global War on Terrorism, the relevance of biometric technology has grown exponentially. The military must achieve identity dominance, where U.S. military forces have the distinct ability to separate friend from foe by linking people to their previous identities and past terrorist or criminal activities. We can use biometric technology to achieve identity dominance and must deploy it to meet the requirements of force protection, actionable intelligence, and law enforcement. Establishing identity dominance through a comprehensive Automated Biometric Identification System (ABIS) will enable the U.S. military to identify friend or foe to keep America safer., Originally published in Militart Review, p30-34, Sep/Oct 2005. This product is part of the RAND Corporation reprint series. RAND reprints reproduce previously published journal articles and book chapters with the permission of the publisher. RAND reprints have been formally reviewed in accordance with the publisher's editorial policy.

This paper describes a prototype system for the intelligent classification of electrolaryngograph (EGG) signals in order to provide an objective assessment of voice quality in patients at different stages of recovery after treatment for larynx cancer. The system extracts salient short-term and long-term time-domain and frequency-domain parameters from EGG signals taken from male patients steadily phonating the vowel /i/. The quality of these voices was also independently assessed by a Speech and Language Therapist (SALT) according to their 7-point ranking of subjective voice quality. These data were used to train and test a Multi-layer Perceptron (MLP) neural network to classify EGG signals in terms of voice quality. Several MLP configurations were investigated using various combinations of these signal parameters, and the best results were obtained using a combination of short-term and long-term parameters, for which an accuracy of 92% was achieved. It is envisaged that this system could be used as a valuable aid to the SALT during clinical evaluation of voice quality., Papers from the 23rd Annual International conference of the IEEE Engineering in Medicine and Biology Society, October 25-28, 2001, held in Istanbul, Turkey. See also ADM001351 for entire conference on cd-rom. The original document contains color images.

A major application of Speaker Identification (SI) is suspect identification by voice. This thesis investigates techniques that can be used to improve SI technology as applied to suspect identification. Speech Coding techniques have become integrated into many of our modern voice communications systems. This prompts the question - how are automatic speaker identification systems and modern forensic identification techniques affected by the introduction of digitally coded speech channels? Presented in this thesis are three separate studies investigating the effects of speech coding and compression on current speaker recognition techniques. A relatively new Spectral Analysis technique - Higher Order Spectral Analysis (HOSA) - has been identified as a potential candidate for improving some aspects of forensic speaker identification tasks. Presented in this thesis is a study investigating the application of HOSA to improve the robustness of current ASR techniques in the presence of additive Gaussian noise. Results from our investigations reveal that incremental improvements in each of these aspects related to automatic and forensic identification are achievable.

This thesis investigates aspects of encoding the speech spectrum at low bit rates, with extensions to the effect of such coding on automatic speaker identification. Vector quantization (VQ) is a technique for jointly quantizing a block of samples at once, in order to reduce the bit rate of a coding system. The major drawback in using VQ is the complexity of the encoder. Recent research has indicated the potential applicability of the VQ method to speech when product code vector quantization (PCVQ) techniques are utilized. The focus of this research is the efficient representation, calculation and utilization of the speech model as stored in the PCVQ codebook. In this thesis, several VQ approaches are evaluated, and the efficacy of two training algorithms is compared experimentally. It is then shown that these productcode vector quantization algorithms may be augmented with lossless compression algorithms, thus yielding an improved overall compression rate. An approach using a statistical model for the vector codebook indices for subsequent lossless compression is introduced. This coupling of lossy compression and lossless compression enables further compression gain. It is demonstrated that this approach is able to reduce the bit rate requirement from the current 24 bits per 20 millisecond frame to below 20, using a standard spectral distortion metric for comparison. Several fast-search VQ methods for use in speech spectrum coding have been evaluated. The usefulness of fast-search algorithms is highly dependent upon the source characteristics and, although previous research has been undertaken for coding of images using VQ codebooks trained with the source samples directly, the product-code structured codebooks for speech spectrum quantization place new constraints on the search methodology. The second major focus of the research is an investigation of the effect of lowrate spectral compression methods on the task of automatic speaker identification. The motivation

The Multilingual Interview System is a Windows-based application program designed to let users conduct simple interviews by voice in languages they do not speak. Any statement or question that is within the vocabulary, when spoken into a microphone attached to the computer, is recognized by a large-vocabulary speech recognition system and converted into a sequence of pre-recorded wave files which are they played back through a loudspeaker attached to the computer. The development of the operational applications for this system was done by NOMI in Pensacola, Florida. Dragon Systems, Inc. developed and customized the underlying speech recognition technology and produced the application software that connects the spoken input to the spoken output as well as the software to record foreign-language wave files for playback. The system was tested at Ft. Bragg and deployed in Bosnia on laptop, handheld and wearable computers.

Speaker recognition, like other biometric personal identification techniques, depends upon a person's intrinsic characteristics. A realistically viable system must be capable of dealing with the open-set task. This effort attacks the open-set task, identifying the best features to use, and proposes the use of a fuzzy classifier followed by hypothesis testing as a model for text-independent, open-set speaker recognition. Using the TIMIT corpus and Rome Laboratory's GREENFLAG tactical communications corpus, this thesis demonstrates that the proposed system succeeded in open-set speaker recognition. Considering the fact that extremely short utterances were used to train the system (compared to other closed-set speaker identification work), this system attained reasonable open-set classification error rates as low as 23% for TIMIT and 26% for GREENFLAG. Feature analysis identified the liftered linear prediction cepstral coefficients with or without the normalized log energy or pitch appended as a robust feature set (based on the 17 feature sets considered), well suited for clean speech and speech degraded by tactical communications channels.

One of the main input devices to computerized systems is Voice Recognition Systems (VRS). VRS is best suited for applications where job functions require more than two hands to be performed. The performance of VRS is highly dependent on the environment's noise. The higher the noise level the lower the recognition capability. Automatic lip reading through vision systems have been utilized to improve the recognition capability in noisy environments. However, this approach is costly and time-consuming. The objective of this thesis was to investigate the utilization of an Infrared sensor for automatic lip reading to improve the recognition performance of VRS. The developed system is cheaper and faster than other automatic lip readers. The test results of fifty words and eleven digits indicated that the method has good repeatability, and good character recognition, while not dependent on or sensitive to the ambient light level. Although speaker independence was tested, the results are inconclusive.

The Remote Medical Diagnosis System (RMDS) is being developed to link ships and shore stations lacking professional medical staffs with larger ships and stations having such staffs. Although RMDS will be capable of transmitting a variety of visual and auditory medical data (eg, X-rays, ECGs, voice and heart- lung sounds, etc), its most important use will be for the transmission of high- quality X-ray images. Because of system limitations and other design factors, the resolution of the system must be limited to no more than that necessary for functionally satisfactory performance. As used here, functionally satisfactory performance specifically means that the performance of the radiologist in interpreting X-rays and related data transmitted by RMDS must be substantially as effective as his performance in interpreting original X-ray images. In order to determine the resolution which will actually be required in use of the system, a series of tests was planned and conducted. In these tests, radiologists at the San Diego Naval Regional Medical Center interpreted test X- rays which were displayed at various levels of resolution. The responses of the radiologists were analyzed to derive the minimum resolution levels which would be compatible with system design and operational requirements. The body of the report describes the experiment, the results obtained, and the conclusions drawn from the results including specific guidelines for resolution needed in the RMDS visual channel.

The purpose of this experiment was to determine if the high accuracy rate of current voice recognitiion systems would be reduced significantly if speakers were required to enter utterances through a mask, as opposed to the boom microphone used with most conventional voice recognition systems. The findings suggest that no significant increase in non-recognitions (e.g., errors where the system rejects the input and says, in effect, (I don't understand you, say it again) is evident while using a mask. Misrecognitions (i.e., errors where the system accepts the input but mistakes it for a different input) do increase significantly under masked conditions. However, the data also indicate that prior experience with speaking into masks or microphones may be a significant moderator of this relationship; subjects that reported having had little or no experience speaking into masks or microphones showed significantly more misrecognition errors than those that reported having some or considerable experience speaking into masks or microphones. Moreover, the data indicate that, when using masks, those subjects that reported having had experience with speaking into masks and microphones (e.g., pilots, communicators) displayed misrecognition error rates still statistically different from but much more comparable to the error rates displayed by subjects under no-mask conditions.

This research sought to isolate physiological attributes which would be of use in automatic identity verification for access control. The work proceeded in three stages. During the initial phase, over 30 potential physical attributes were evaluated through mathematical modeling, opinion survey, and system analysis. The two best attributes, handprints and a modified electrocardiogram called a 'C-trace'--were selected for in depth study. Data on both attributes was collected on 72 subjects at 3 sessions over 3 months during the 2nd phase. In the 3rd phase, the data was analyzed to determine the type i and ii error rates for identity verification. A type I error is a failure to recognize a legitimate subject, whereas a type II error occurs when an imposter is identified as a legitimate subject. Results were quite encouraging. The handprint achieved a type I error %of 1.4% and a type II of 1.6%. The c-trace performed at a type I of 1.2% and a type II of 1.1%. Arguments which indicate that the handprint results can be improved significantly and the c-trace moderately are given.

Speed and accuracy have been compared for isolated-word voice recognition, keyboard, and graphical menu data entry systems. One entry task involved simple copying of numeric and alphanumeric data strings. A second was a simulation of complex flight data entry scenario. The factors evaluated included voice response feedback and prompting, hand occupation during data entry, and subject experience. Keyboard provided the fastest and most accurate entry of numeric data strings and the fastest entry of alphanumeric strings by subjects with keyboard experience in the simple scenario, but was slow relative to voice and graphical menu for entry of words by inexperienced subjects in the complex scenario. Voice entry provided the lowest error rate for entry of alphanumeric data strings in the simple scenario primarily because of its greater immunity to reading errors. In the complex scenario, voice was faster than keyboard for inexperienced subjects, and had a similar operational error rate, but had a substantially higher error rate before correction. Graphical menu ranked between keyboard and voice in most of the simple scenario measures, except that it was least accurate with alphanumeric data, and had the lowest entry speed for long strings.

The effects of g-force stress on human voice patterns were investigated with the objective of finding means for making isolated word recognition word devices work in the fighter aircraft cockpit environment. Data were taken in a human centrifuge with SCOPE Electronics Inc's Voice Data Entry System (VDETS) used to prompt and pace the subjects. Data were subsequently digitized and stored for analysis and recognition experiments using the VDETS algorithm with a number of variations. Recognition performance on the centrifuge data was initially poor. Means were found for improving it substantially through modifications to the VDETS algorithm and through preprocessing techniques. VDETS modifications included increased coding resolution, improved segmentation techniques and provision for multimode training. Breathing noise elimination and inverse filtering preprocessing routines were effective. Variations in spectral characteristics with g-force stress were found, but no consistent pattern was discerned. The effectiveness of the inverse filtering led to the conclusion that the major problem was the face mask worn by the subjects, causing a variable element in the acoustic transmission path. Additional work will be required to eliminate face mask effects. (Author)

This project was initiated to test the feasibility of designing a close-coupled, two-way communication link between man and computer using biological information. Specifically, experiments have been devised to determine whether biological information can be related to human thought, whether this information can be processed meaningfully by a computer, and whether similar biological processes representing the same or other thoughts can be induced in the same or another individual. Should such a close-coupling between man and machine prove to be feasible, possible applications would include extremely rapid interactive processing between a man and a computer, or communication between two or more persons with the computer acting as an interface.