Performance Improvements of the Phoneme Recognition Algorithm.
Abstract
An analysis of speech is made by comparing single, two and three time slice phonemes to five time slice phonemes (.08 sec/time slice). All phonemes created from the same speech with the single, two and three time slice phonemes created from portions of the five time slice phonemes. It was found that the single time slice phonemes compared favorably with the five time slice phonemes in recognizing speech only if the frequency components remained relatively constant over a period of time, such as the vowel and nasal sounds. The two time slice phonemes showed results that began to duplicate those of the five time slice phonemes, but still had inconsistent results identifying fricative sounds. Three time slice phonemes results showed a closer correlation with the results of the five time slice phonemes than those of the one and two time slice phonemes. All results were obtained using a 64 point sampled, Hamming windowed, Discrete Fourier Transform. The recognition results for each time slice of speech, using various length phonemes, are tabulated and the results are used to re-synthesize the original speech. This was done by using digitized speech composed of the middle time slices from the 71 five time slice phonemes. Results indicated that the synthesized speech was understandable when the recognition results successfully identified the proper phoneme for approximately 4 consecutive time slices. An extraneous phoneme choice in a consecutive grouping of a phoneme choice did not seriously degrade the output since it accounted for only an .08 second time slice. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 1984
- Accession Number
- ADA144562
Entities
People
- J. E. Fletcher
Organizations
- Air Force Institute of Technology