Marquette Home | College of Engineering

Home

People

Publications

Labs

Bibliography

Links

Downloads

**New! RPS and TKDE Matlab toolboxs now available!**

Project Overview

Summary

Current state-of-the-art speech recognition systems generally use Hidden Markov Models (HMMs) with frame-based spectral measures (often cepstral coefficients) as the primary features. Traditional spectral analysis techniques have been used for many years, with progress in recognition accuracy over the last 10-15 years being primarily incremental. This research project focuses on the development of a significantly different approach to characterizing speech signals, based on state-of-the-art techniques for time-series modeling. These time-series techniques combine state-space embedding methods and learning algorithms to create highly accurate non-linear models of a system's state. This research integrates a dynamical systems approach with a continuous speech recognition system, changing the analytical focus from the frequency domain to the time domain. The time-delay embedding technique, taken from dynamical systems theory, is used to reconstruct the state spaces of the speech waveforms. The resulting state spaces are then characterized to generate a set of features, which are evaluated with respect to their ability to differentiate the individual phonemes that are the building blocks of speech.

Objectives

The focus of this project is to use time-domain analysis of speech to create new modeling techniques and to gain a better understanding of speech signals, leading to a subsequent improvement in speech recognition accuracy. To achieve this, the primary research objectives include the application of the time domain embedding approach to the characterization of speech signals, the development of an effective model for measuring differences between the signals, and the integration of this model with an HMM-based speech recognition system. The speech tasks used for implementation of these objectives include both isolated phoneme recognition and continuous word recognition experiments.

Methods

Successful achievement of the above objectives requires the development of several new technologies. For the characterization of speech signals in the time domain, the Time Series Data Mining approach, which has been successfully applied to event prediction, is modified for application to speech waveforms, including the development of techniques for identifying optimal lag times for the time-domain embedding process. Stochastic methods, including various clustering techniques for learning parametric densities such as Gaussian Mixture Models, are used for identifying appropriate feature representations of the embedded waveforms. For integrating these features with a recognition system, an HMM-based speech system is modified to use the new time-domain features for computing state occupancy likelihoods within the training and recognition algorithms.

Impact

The impact of these new technologies and their application to the speech recognition task extends into both the machine learning and signal processing communities. The development of time-domain characterization methods is directly applicable to many problems of interest in the chaos and non-linear modeling domains. These new methods are able to concretely measure differences between the phase-space representations of dynamical systems. The application to the speech recognition task is particularly appropriate for this research, since it is a novel approach in a field where traditional linear systems approaches have been unable to achieve fully satisfactory results. It is expected that the experiments conducted will lead to significant gains with respect to a fundamental understanding of the characteristics and analysis of speech signals, with potential long-term application to other areas of speech processing such as speech coding and synthesis.

Demonstration videos (in AVI format)

Video 1
Video 2
Video 3
Video 4

Contact information


Knowledge and Information Discovery Laboratory
Olin Hall of Engineering
Marquette University
P.O. 1881
Milwaukee, Wi 53201-1881

Faculty offices: Olin 523, (414) 288-6046, Haggerty 224, (414) 288-7088
Grad student offices: Olin 523, (414) 288-6046
Computer and research lab: Olin 523, (414) 288-3503

Speech and Signal Processing Laboratory
Olin Hall of Engineering
Marquette University
P.O. 1881
Milwaukee, WI 53201-1881

Faculty offices: Olin 518D, (414) 288-1608, Haggerty 214, (414) 288-0631
Grad student offices: Olin 518, (414) 288-7451
Computer and research lab: Olin 518A, (414) 288-3503
Data collection lab: Olin 518B, (414) 288-3503

Faculty contacts:
Michael T. Johnson, Ph.d.
web: http://www.eng.mu.edu/johnson
email: mike.johnson@marquette.edu


Richard J. Povinelli, Ph.d.
web: http://povinelli.eece.mu.edu
email: richard.povinelli@marquette.edu

 

 

< Speech Recognition Project Home


This material is based upon work supported by the National Science
Foundation under Grant No. 0113508.

Any opinions, findings, and conclusions or recommendations expressed
in this material are those of the author(s) and do not necessarily
reflect the views of the National Science Foundation

©2001-2002 Michael T. Johnson & Richard J. Povinelli -- Last Update: July 25, 2003