Collingham, Russell James (1994) Towards an automatic speech recognition system for use by deaf students in lectures. Unspecified thesis, Durham University.
According to the Royal National Institute for Deaf people there are nearly 7.5 million hearing-impaired people in Great Britain. Human-operated machine transcription systems, such as Palantype, achieve low word error rates in real-time. The disadvantage is that they are very expensive to use because of the difficulty in training operators, making them impractical for everyday use in higher education. Existing automatic speech recognition systems also achieve low word error rates, the disadvantages being that they work for read speech in a restricted domain. Moving a system to a new domain requires a large amount of relevant data, for training acoustic and language models. The adopted solution makes use of an existing continuous speech phoneme recognition system as a front-end to a word recognition sub-system. The subsystem generates a lattice of word hypotheses using dynamic programming with robust parameter estimation obtained using evolutionary programming. Sentence hypotheses are obtained by parsing the word lattice using a beam search and contributing knowledge consisting of anti-grammar rules, that check the syntactic incorrectness’ of word sequences, and word frequency information. On an unseen spontaneous lecture taken from the Lund Corpus and using a dictionary containing "2637 words, the system achieved 815% words correct with 15% simulated phoneme error, and 73.1% words correct with 25% simulated phoneme error. The system was also evaluated on 113 Wall Street Journal sentences. The achievements of the work are a domain independent method, using the anti- grammar, to reduce the word lattice search space whilst allowing normal spontaneous English to be spoken; a system designed to allow integration with new sources of knowledge, such as semantics or prosody, providing a test-bench for determining the impact of different knowledge upon word lattice parsing without the need for the underlying speech recognition hardware; the robustness of the word lattice generation using parameters that withstand changes in vocabulary and domain.
|Item Type:||Thesis (Unspecified)|
|Copyright:||Copyright of this thesis is held by the author|
|Deposited On:||16 Nov 2012 11:00|