Speech Database Development
Abstract
The development of an acoustic - phonetic database is thought to be crucial to the speech program because the acoustic realization of phonemes depends on complex interactions among a multitude of factors. Therefore, in order to successfully develop a speaker-independent, phonetically-based speech recognition system, a large body of speech data, collected from many speakers, is needed to help us discover and quantify these context-dependent phenomena. In addition, the speech database can serve two other functions. First, it can be used for training certain speech recognition systems. For some algorithms, such as hidden Markov modelling (HMM), a large amount of training data is needed to obtain stable estimates of the parameters of the stochastic models. For rule- based algorithms, substantial amounts of data are also needed in order to set proper thresholds on speech parameters. Second, the database can be used for performance evaluation. Given the many different approaches to the speech recognition problem, it is often difficult to compare their relative merits. Testing specific recognition algorithms or entire speech recognition systems on a common database will provide a means to evaluate their relative performance. Keywords: Systems engineering; Systems analysis; Speech communications.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 21, 1988
- Accession Number
- ADA202461
Entities
People
- Victor W. Zue
Organizations
- Massachusetts Institute of Technology