Limited Continuous Speech Recognition by Phoneme Analysis.
Abstract
A Limited Continuous Speech Recognition system is developed based upon phoneme analysis. 16 Bandpass filters are used to obtain the frequency components of the input speech. The input speech is broken into packets of 40 milliseconds each. These packets are compared with phonemes in a template file by a differencing of frequency magnitudes. The resulting phoneme string representation of the input speech is compressed and compared with strings in a library file for discrete word recognition. For continuous speech recognition the phoneme string is analyzed a phoneme at a time to construct word sequences. The word string which best matches the input phoneme string is recognized as the word sequence. The system has an accuracy of about 94% for discrete word recognition and about 80% for continuous speech recognition. The vocabulary used is the digits zero to nine and point. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 08, 1983
- Accession Number
- ADA138021
Entities
People
- A. Hussain
Organizations
- Air Force Institute of Technology