Z-34 Device Kai Lassfolk; Sound Processing Kit-User's Guide; SPKit WaveShaper-User's Guide; 1995; 5 pages. Marc Le Brun; Journal of the Audio Engineering Society; Digital Waveshaping Synthesis; Apr. 1, 1979; 16 pages. Daniel Arfib; Journal of the Audio Engineering Society; Digital Synthesis of Complex Spectra by Means of Multi plication of Nonlinear Distorted Sine Waves; Oct. 1, 1979; 12 pages. Hal Chamberlin; Musical Applications of Microproces sors-Second Edition; pp. 48-54, 473–480; 1985; 15 pages. Ronald E. Crochiere; Lawrence R. Rabiner; Multirate Digi tal Signal Processing; 1983; reference text.

This paper describes a set of 58 Matlab®-based speech processing exercises designed to give students and instructors hands-on experience with digital speech processing basics, fundamentals, representations, algorithms and applications. This result is achieved by providing working Matlab code using a LITE graphical user interface (GUI) for ease of use and understanding of the code. For each Matlab exercise we provide a brief description of the code and GUI structure, an explanation of key technical aspects of the solution, a description of the organization of the Lite Graphical User Interface, a scripted run of the software with an extensive set of speech files for processing, some results figures generated by the provided code, and a set of technical issues for further experimentation with the code. Our goal with these exercises is to provide a set of speech processing tools for the instructor and an extended set of speech processing exercises (denoted as “Issues for Further Experimentation”) that students can use to develop their understanding of the technical content of each speech processing exercise. This paper describes the resulting set of exercises and shows how they can be used to enhance the teaching and learning of digital speech processing.

We have developed a novel loss function that embeds large-margin classification into Minimum Classification Error (MCE) training. Unlike previous efforts this approach employs a loss function that is bounded, does not require incremental adjustment of the margin or prior MCE training. It extends the Bayes risk formulation of MCE using Parzen Window estimation to incorporate large-margin classification and develops a loss function that is a sum of shifted sigmoids. Experimental results show improvement in recognition performance when evaluated on the TIDigits database.

An automatic speech recognition system has two main components, the front end feature processing component, followed by a model training component. A widely used algorithm for training models is Minimum Classification Error (MCE) that designs the model parameters to minimize recognition error. One approach for designing the feature processing component to minimize the error in recognition is transforming the features using the MCE criterion that is most commonly used only to estimate the model parameters. Past efforts that have integrated feature transformation into MCE training assumed the Hidden Markov Model state distributions were represented by diagonal covariance Gaussian mixtures when estimating the model parameters; but assumed full covariance Gaussian mixtures when estimating the feature transformation matrix. We rectify this discrepancy in assumptions, derive and implement the MCE based feature transformation, using diagonal covariance Gaussian mixtures. For designing the model parameters, MCE minimizes the recognition error using a standard sigmoid loss function, which is a Parzen window estimate of the Bayes risk. Using different kernels for Parzen Window estimation we developed new loss functions, viz. the Gaussian Kernel based loss and Generalized Savage loss. Investigation into the recognition performance of these loss functions lead to the introduction of a new large margin based loss function for MCE (LM-MCE) where the error is minimized and margin maximized. Minimizing the error, aims to shift the decision boundary so the wrongly classified tokens move to the side of the correct class, (however some of these tokens could lie on or close to the boundary); while increasing the margin between the correctly classified tokens and the decision boundary, removes some of these tokens away from the boundary and improves the robustness of the classification. Unlike previous studies, this effort does not require MCE training prior to maximizing the margin nor does it require non-optimal manual increments of the margin shifts; and since this loss function is bounded, it is not susceptible to outliers. The development of the new LM-MCE loss function has a theoretical basis in the Vapnik and Chervonenkis (VC) theory and was developed using the Bayes risk formulation.

A new loss function has been introduced for Minimum Classification Error, that approaches optimal Bayes’ risk and also gives an improvement in performance over standard MCE systems when evaluated on the Aurora connected digits database.

The proliferation of network-connected media-enabled devices has given users access to large volumes of information and entertainment in video form. Taking advantage of these vast video resources involves the creation of effective mechanisms for searching, navigating, personalizing, and repurposing video to support alternative consumption modes. Automated content analysis algorithms that utilize media processing techniques are the key to the creation of such mechanisms. Media processing also serves to facilitate retrieval and navigation of content by enabling multimodal user interfaces. In this talk I will discuss some of the media processing research at AT&T Labs aimed at extracting content-based metadata, identifying relevant segments to create short-form personalized content, and providing speech-enabled multimodal user interfaces. I will describe several prototype systems based on these capabilities for giving users easy access to video and multimedia information on a wide range of media-enabled devices.

A Minimum Classification Error (MCE) based recognition system that also estimates a global feature transformation matrix has been implemented. Unlike earlier studies, we make the explicit assumption that the covariance matrix of the Gaussian mixtures is diagonal when estimating the transformation matrix. This is necessary for mathematical consistency between the model and the transformation matrix estimates. Experimental results show a reduction of up to 50% in the word error rate as compared to Maximum Likelihood estimation.

Statistical data-driven methods and knowledge-based methods are two recent trends in Automatic Speech Recognition (ASR). Hidden Markov Model (HMM)-based speech recognition techniques have achieved great success for controlled tasks and environments. However, when we require improved accuracy and robustness (closer to Human Speech Recognition (HSR)), HMM algorithms for speech recognition gradually fail. Hence a need has emerged to incorporate higher level linguistic information into ASR systems in order to further discriminate between speech classes or phonemes with high confusion rates. The Automatic Speech Attribute Transcription (ASAT) project is one of the recent research efforts that has tried to bridge the gap between ASR and HSR. In this thesis we focus on the design and optimization of the front end processing of the ASAT system, whose goal is to estimate a set of attribute and phoneme probability lattices which can be combined with information from higher level knowledge sources in a set of speech event verification modules in order to make a final recognition decision. We propose a set of both frame-based methods and segment-based methods to improve the recognition performance of distinctive features and phonemes in English. We also study and evaluate both a parallel speech feature organization and a hierarchical phoneme topology. There are 4 main parts in this thesis work. In the first part, we use frame-based methods to estimate the likelihood of static sounds (e.g., steady vowels, fricatives, etc), and implement the parallel feature detection using Multi-Layer Perceptrons (MLPs) in order to detect the 14 Sound Pattern of English (SPE) features. In the second part, we use segment-based methods to classify dynamic sounds (e.g., stop consonants, diphthongs, etc), and use Time-Delay Neural Networks (TDNNs) to recognize phoneme classes in a hierarchical phoneme and feature organization. In the third part and in the forth part, we combine the frame-based parallel speech feature detection system and the segment-based hierarchical phoneme classification system to improve the overall phoneme classification performance and the speech feature detection performance. The main contribution of this thesis is the creation of a phoneme recognizer that overcomes the disadvantages of pure statistical or knowledge-based systems, and provides a way to incorporate acoustic/phonetic/linguistic knowledge into an existing (HMM-based) automatic speech recognition system.

Phonemes in the English language can be represented using either parallel or hierarchical distinctive speech features. There have been a number of efforts to integrate multiple information sources but none of these efforts addressed the issue of combining multiple sets of articulatory/linguistic features with different organization topologies. In this study, we combine a frame-based parallel speech feature detection system and a segment-based hierarchical phoneme classification system with the goal of improving the overall phoneme classification performance. We describe a mathematical framework for a unified classification system in which frame-based parallel feature detection is incorporated into a segment-based hierarchical phoneme classification method. Experimental results show that the combined system provides improved classification performance for some phoneme classes.

Time-Delay Neural Networks (TDNN) have been shown by Waibel et al. [1] to be a good method for the classification of dynamic speech sounds such as voiced stop consonants. In this paper we discuss key issues in the design and training of a TDNN, based on a Multi-Layer Perceptron (MLP), when used for classification of the sets of voiced stop consonants (/b/, /d/, and /g/) and unvoiced stop consonants (/p/, /t/ and /k/) from the TIMIT database. We show that by transforming each input parameter to the TDNN to be a zero mean, unit variance distribution (separately for each phoneme class) we can greatly improve the overall classification performance. The resulting TDNN classification accuracy for voiced or unvoiced stop consonants is around 91%. This performance is achieved without any specific discriminative spectral measurements and can be applied directly to the classification of any of the dynamic phoneme classes.

Automatic Speech Attribute Transcription (ASAT), an ITR project sponsored under the NSF grant (IIS-04-27113), is a cross-institute effort involving Georgia Institute of Technology, The Ohio State University, University of California at Berkeley, and Rutgers University. This project approaches speech recognition from a more linguistic perspective: unlike traditional ASR systems, humans detect acoustic and auditory cues, weigh and combine them to form theories, and then process these cognitive hypotheses until linguistically and pragmatically consistent speech understanding is achieved. A major goal of the ASAT paradigm is to develop a detection-based approach to automatic speech recognition (ASR) based on attribute detection and knowledge integration. We report on progress of the ASAT project, present a sharable platform for community collaboration, and highlight areas of potential interdisciplinary ASR research. Index Terms: attributes, events, features, detection, speech recognition, speech attribute transcription, utterance verification

Since even before the time of Alexander Graham Bell's revolutionary invention, engineers and scientists have studied the phenomenon of speech communication with an eye on creating more efficient and effective systems of human-to-human and human-to-machine communication. Starting in the 1960s, digital signal processing (DSP), assumed a central role in speech studies, and today DSP is the key to realizing the fruits of the knowledge that has been gained through decades of research. Concomitant advances in integrated circuit technology and computer architecture have aligned to create a technological environment with virtually limitless opportunities for innovation in speech communication applications. In this text, we highlight the central role of DSP techniques in modern speech communication research and applications. We present a comprehensive overview of digital speech processing that ranges from the basic nature of the speech signal, through a variety of methods of representing speech in digital form, to applications in voice communication and automatic synthesis and recognition of speech. The breadth of this subject does not allow us to discuss any aspect of speech processing to great depth; hence our goal is to provide a useful introduction to the wide range of important concepts that comprise the field of digital speech processing. A more comprehensive treatment will appear in the forthcoming book, Theory and Application of Digital Speech Processing [101].

Camera calibration is an important step in 3D reconstruction of scenes. Many natural and man made objects are circular and form good candidates as calibration objects. We present a linear calibration algorithm to estimate the intrinsic camera parameters using at least three images of concentric circles of unknown radii. Novel methods to determine the projected center of concentric circles of unknown radii using the projective invariant, cross ratio, and calculating the vanishing line of the circle are proposed. The circular calibration pattern can be easily and accurately created. The calibration algorithm does not require any measurements of the scene or the homography between the images. Once the camera is fully calibrated the focal length of zooming cameras can be estimated from a single image. The algorithm was tested with real and synthetic images with different noise levels.

Earlier research has shown that the maximum spectral transition positions are related with the perceptual critical points that contain the most important information for consonant and syllable perception. This paper presents a quantitative analysis of the relation, in time, between the maximum spectral transition positions and the phone boundaries in fluent read speech. This analysis is based on the training part of the TIMIT American English database which contains both phone boundaries and labels manually-determined by a group of experts. The results of this analysis show that there is a significant correlation between the maximum spectral transition positions and the manually selected phone boundaries. This, in turn, suggests that there is an important relation between the commonly accepted phone boundaries and the perceptual critical points. Index Terms: phone boundaries, spectral transition, phonemes

Automatic chord recognition has been a topic of interest in the context of Music Information Retrieval (MIR) for several years, and attempts have been made at implementing such systems using well understood Signal Processing and Pattern Recognition techniques. The sequence of chords in a musical recording, in addition to providing the melody, often provides the most effective way to summarize a piece of music. Information on the chord structure of a musical work is also useful in the context of Interactive Music Systems. In nearly every previous chord recognition system, a summation of filterbank outputs, called the Pitch Class Profile (PCP), (which is a measure of the spectral energy in different frequency regions), has been used to represent the signal. Because spectral characteristics of the musical sounds vary between different sound sources,spectral energy measurements (if not properly preprocessed)will capture information inherent to a particular instrument.In this paper we give experimental results that show the benefits of proper signal processing before calculating the spectral energy via the PCP (or any related measure). Also,we discuss a method for finding the boundaries between different chord labels by suitably combining multiple measures of audio novelty. Together, the segmentation and classification methods that we have developed have enabled high recognition rates for musical chords.