Related Experiment Videos
Considerations in applying clustering techniques to speaker-independent word recognition
The Journal of the Acoustical Society of America
|September 1, 1979
Summary
This study explores automatic methods for creating speaker-independent word templates for speech recognition. The first method, using distance data for clustering, proved superior for robust word recognition systems.
Area of Science:
- Speech Recognition
- Pattern Recognition
- Computational Linguistics
Background:
- Previous research demonstrated the effectiveness of experimenter-guided clustering for creating speaker-independent word templates.
- Successful template sets were representative of a large talker population in prior studies.
Purpose of the Study:
- To investigate fully automatic techniques for clustering word variations into speaker-independent templates.
- To compare the efficacy of two novel automatic clustering methods for isolated word recognition.
Main Methods:
- Method 1: Utilizes distance data between word replications to form clusters, with templates derived from cluster minimax or averages.
- Method 2: A variation of Rabiner's approach, combining averaging and nearest neighbor rules for simultaneous template and cluster definition.
Main Results:
- Experimental data indicate that the first method (distance-based clustering) outperforms the second method.
- The superiority of the first method was observed when employing three or more clusters per word.
Conclusions:
- Automatic clustering techniques are viable for generating speaker-independent word templates.
- Distance-based segmentation into clusters offers a more effective approach for isolated word recognition compared to the combined averaging and nearest neighbor rule method, particularly with multiple clusters.