Linguistic knowledge for Speech recognition By Ahmed Aly

Linguistic knowledge for Speech recognition By : Ahmed Aly 06/05/2013

Project description The main goal of this project is to study the effect of using linguistics knowledge on the task of speech recognition. I am studying the usage of such knowledge in the following contexts : Using higher level linguistics knowledge for speech recognition error correction Using prosody models for conversational speech recognition The effect of using syntactic and semantic information on the performance of speech recognition error detection The effect of using prosody features in spoken speech subjectivity analysis

Work done so far … Studying theoretical part of the automatic speech recognizer module Literature review of the usage of linguistics knowledge in the speech recognition task Literature review of the usage of prosody in spoken speech subjectivity analysis

Main findings so far … In the context of error correction It has been shown that the usage of syllablebased models has shown s a superior performance in domain-specific IR applications in comparison to word-based models In this case a relatively smaller training set is needed. And it could handle intraword transformation and syllable-to-syllable transformation The usage of semantic knowledge has proved also to be successful in the task of error correction. An example of this semantic knowledge is lexico-semantic pattern (LSP) which is a structure where linguistic entries and semantic types are used in combination to abstract certain sequences of the words in a text

Main findings so far … (continued) In the context of prosody experiments in acoustic model clustering show that representing syllable position and stress as conditioning factors lead to gains in recognition performance and reduced system complexity Other experiments with pronunciation modeling show gains in surface-form phone prediction due to directly conditioning on acoustic-prosodic features For subjectivity analysis, It has been shown that the usage of prosody features is not very usefull as Ngram characters and words features outperformed prosody features

Q&A