A LIMITED-VOCABULARY, MULTI-SPEAKER AUTOMATIC ISOLATED WORD RECOGNITION SYSTEM,
Abstract
Techniques for automatic recognition of isolated words are investigated, and a computer simulation of a word recognition system is effected. The following aspects of the system are considered in detail, data acquisition and digitizing, word detection, amplitude and time normalization, short-time spectral estimation including spectral windowing, spectral envelope approximation, parametric and adaptive feature extraction, and pattern classification. A method of time normalization is developed and is shown very effective in reducing recognition error rates. Recognition parameters are derived from estimated short-time power spectra. An evaluation of spectral windows is made resulting in the selection of the hanning window. Fourier transformation of finite length time series is discussed along with the modified periodogram. A convolutional smoothing method of spectral envelope approximation is developed, and several smoothing windows are compared. Features are extracted from partitioned intervals over the spectral envelopes. The boundaries of these intervals are determined using both parametric and adaptive procedures. Formant orientation is found effective for the parametric boundaries. An improved measure of feature merit is developed and is compared with both the information and analysis-of-variance criteria. The overall word recognition system was found to perform very satisfactorily with recognition accuracies reaching 99 percent for a single speaker, 50-word vocabulary and 96 percent for a three-speaker, 30-word vocabulary. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 1969
- Accession Number
- AD0696206
Entities
People
- James E. Paul Jr
Organizations
- North Carolina State University