Related Experiment Videos
Structural design of hidden Markov model speech recognizer using multivalued phonetic features: comparison with
1Department of Electrical and Computer Engineering, University of Waterloo, Ontario, Canada.
The Journal of the Acoustical Society of America
|December 1, 1992
Summary
This study introduces a new phonetic-feature-based speech recognition system using hidden Markov models (HMMs). This novel approach significantly reduces error rates compared to traditional methods for speaker-dependent tasks.
Area of Science:
- Speech Recognition
- Acoustics
- Machine Learning
Background:
- Conventional speech recognition methods often struggle with coarticulation and data sharing.
- Utilizing phonetic features as basic speech units offers a promising alternative.
Purpose of the Study:
- To present and evaluate a novel speech recognition approach based on multidimensional phonetic features.
- To demonstrate the effectiveness of this feature-based system in reducing error rates.
Main Methods:
- Employed a hidden Markov model (HMM) framework to track the temporal evolution of phonetic features.
- Designed a recognizer state topology guided by speech knowledge for a stop consonant-vowel vocabulary.
- Evaluated speaker-dependent stop consonant discrimination using data from 15 speakers.
Main Results:
- The feature-based recognizer effectively accommodates coarticulatory effects like feature spreading and formant transitions.
- Achieved significant error rate reductions: 23% vs. words, 37% vs. phonemes, 42% vs. allophones, and 38% vs. microsegments.
- Demonstrated high acoustic data sharing, making effective use of training data.
Conclusions:
- Phonetic feature-based speech recognition offers a robust framework for improving accuracy.
- This approach provides a significant advantage over conventional HMM-based methods for specific speech tasks.