Although studied for several years, speech recognition is still a field that is developing. Recently several important researchers have pointed out areas within the field that need to be addressed. These include robustness to various environments, large or expandable vocabularies, user-friendliness, high recognition accuracy and the ability to recognise continuous speech. The ability to adapt is an important component of a speech recognition system. People new to the system should have the benefits mentioned above. The system should also manage recognition of different speaking rates. Also, novel environments may cause a drop in the system's performance if it lacks robustness or the ability to adapt. A common target for speech recognition algorithms is to detect the presence of speech units, commonly phonemes. This approach involves grouping speech sounds, or phones, into abstract groups that reflect meaning. Recently artificial neural networks have been applied to this task. Nevertheless, uncertainty and ambiguity are inherent in the neural network recognition process. Several novel techniques are proposed to aid in the recognition process, and to help to fulfil the requirements of a successful speech recognition system. The goal of this research is to investigate theories of speech and language processing that are relevant to speech recognition and spoken language understanding. These theories have their foundations in fields such as engineering, computer science, linguistics, natural language processing, psycholinguistics and psychology.
An adaptive system is implemented to test the validity and usefulness of such work to the fields of speech recognition and spoken language understanding. For example, the development of abstract structures of the human auditory system and the auditory cortex are investigated, and applied towards better engineering methods for building adaptive speech and language systems.
For the implementation of an adaptive speech recognition system, parameters are introduced that can be adjusted either manually or automatically. In this manner, the system can adapt to new speakers and environments. The architecture of the system is modular and hierarchical. Different methods are applied at various levels. For example, artificial neural networks are best suited for low-level processing. A discussion of how errors and uncertainty may be resolved in an unsupervised manner concludes the work. Ideally, the system will adapt to the situation, and the future occurrences of such phenomena may be reduced or eliminated.
