Minimizing Computational Cost for Dynamic Programming Algorithms
Abstract
In this study we introduce and test several methods to reduce the computational cost in dynamic programming algorithms for isolated word recognition systems. Three methods will be discussed in detail: (1) Pruning by preset thresholds, (2) Search based on the Branch and Bound technique, (3) Branch and Bound based search with additional pruning. Compared to conventional algorithms, Method 3 could be seen to yield a speed up of approximately a factor of 5, at no loss of recognition accuracy. The branch and bound method with pruning is also ideally suited for research oriented systems, since pruning is independent of the parametrization used (eliminates the necessity for retuning thresholds). Additional features of this method, which are of importance to maintaining the flexibility and diagnosticity needed for such a system, will be discussed.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 17, 1981
- Accession Number
- ADA104792
Entities
People
- A. Waibel
- N. Krishnan
- Rekha Reddy
Organizations
- Carnegie Mellon University