A comparison of features for large population speaker identification
Master Thesis
2000
Permanent link to this Item
Authors
Supervisors
Journal Title
Link to Journal
Journal ISSN
Volume Title
Publisher
Publisher
University of Cape Town
Department
License
Series
Abstract
Speech recognition systems all have one criterion in common; they perform better in a controlled environment using clean speech. Though performance can be excellent, even exceeding human capabilities for clean speech, systems fail when presented with speech data from more realistic environments such as telephone channels. The differences using a recognizer in clean and noisy environments are extreme, and this causes one of the major obstacles in producing commercial recognition systems to be used in normal environments. It is the lack of performance of speaker recognition systems with telephone channels that this work addresses. The human auditory system is a speech recognizer with excellent performance, especially in noisy environments. Since humans perform well at ignoring noise more than any machine, auditory-based methods are the promising approaches since they attempt to model the working of the human auditory system. These methods have been shown to outperform more conventional signal processing schemes for speech recognition, speech coding, word-recognition and phone classification tasks. Since speaker identification has received lot of attention in speech processing because of its waiting real-world applications, it is attractive to evaluate the performance using auditory models as features. Firstly, this study rums at improving the results for speaker identification. The improvements were made through the use of parameterized feature-sets together with the application of cepstral mean removal for channel equalization. The study is further extended to compare an auditory-based model, the Ensemble Interval Histogram, with mel-scale features, which was shown to perform almost error-free in clean speech. The previous studies of Elli to be more robust to noise were conducted on speaker dependent, small population, isolated words and now are extended to speaker independent, larger population, continuous speech. This study investigates whether the Elli representation is more resistant to telephone noise than mel-cepstrum as was shown in the previous studies, when now for the first time, it is applied for speaker identification task using the state-of-the-art Gaussian mixture model system.
Description
Bibliography: leaves 95-104.
Keywords
Reference:
Baloyi, N. 2000. A comparison of features for large population speaker identification. University of Cape Town.