10-06-2017, 06:52 AM
[attachment=1548]
COMPARISON OF ENERGY-BASED ENDPOINT DETECTORS FOR SPEECH SIGNAL PROCESSING
Accurate endpoint detection is a necessary capability for construction of speech databases from field recordings. In this paper we describe the implementation of two endpoint detection algorithms which use signal features based on energy and rate of zero crossings. We have made extensive use of object-oriented concepts and data-driven programming to make our code re-usable for a variety of applications, including speech recognition. A uniform user-interface for both algorithms has been developed using a novel v i r tual classmethodology.We also present a comparison of the two algorithms using an objective evaluation paradigm we have developed. A small locally prepared database has been used for the purpose of evaluation.
Presented by
A. Ganapathiraju, L. Webster, J. Trimble, K. Bush, P. Kornman
Department of Electrical and Computer Engineering
Mississippi State University
{ganapath,webster,trimble,bush,kornman}@isip.msstate.edu
MOTIVATION FOR AN ENDPOINT DETECTOR
In early stages of speech recognition research, used for the alignment process in DTW isolated word recognizers (ISR) p Still used extensively today as data trimming tools in the preparation of speech databases. l Large databases have become imperative for training the large vocabulary continuous speech recognizers. l Telephone data collection is the easiest and fastest way to collect data for large databases. l A major problem associated with this approach is that the characteristics of speech and noise are not very different in such environments. l To make available as public domain software an accurate and simple tool for endpoint detection
COMPARISON OF ENERGY-BASED ENDPOINT DETECTORS FOR SPEECH SIGNAL PROCESSING
Accurate endpoint detection is a necessary capability for construction of speech databases from field recordings. In this paper we describe the implementation of two endpoint detection algorithms which use signal features based on energy and rate of zero crossings. We have made extensive use of object-oriented concepts and data-driven programming to make our code re-usable for a variety of applications, including speech recognition. A uniform user-interface for both algorithms has been developed using a novel v i r tual classmethodology.We also present a comparison of the two algorithms using an objective evaluation paradigm we have developed. A small locally prepared database has been used for the purpose of evaluation.
Presented by
A. Ganapathiraju, L. Webster, J. Trimble, K. Bush, P. Kornman
Department of Electrical and Computer Engineering
Mississippi State University
{ganapath,webster,trimble,bush,kornman}@isip.msstate.edu
MOTIVATION FOR AN ENDPOINT DETECTOR
In early stages of speech recognition research, used for the alignment process in DTW isolated word recognizers (ISR) p Still used extensively today as data trimming tools in the preparation of speech databases. l Large databases have become imperative for training the large vocabulary continuous speech recognizers. l Telephone data collection is the easiest and fastest way to collect data for large databases. l A major problem associated with this approach is that the characteristics of speech and noise are not very different in such environments. l To make available as public domain software an accurate and simple tool for endpoint detection