ISCA Archive Eurospeech 1999
ISCA Archive Eurospeech 1999

On the use of time alignments for noisy speech recognition

Y. Hauptman, Y. Bistritz

One of the basic aspects of modern pattern matching algorithms used in speech recognition is time-alignment. The use of time-alignment is essential for offsetting speaking rate variations, which is an inherent property of the speech signal. It is known that time-alignment contributes to increased accuracy in speech recognition. However, a key question is whether time-alignment information still contributes to recognition accuracy in highly degraded speech. In this paper we examine the robustness of time-alignment information by introducing a robustness indicator. Isolated words recognition experiments with and without time alignment (using DTW and VQ respectively) are used and to illustrate the issue.