A linguistic study of interview and conversation

The present thesis focuses on the overall structure of the language of two types of Speech Exchange Systems (SES) : Interview (INT) and Conversation (CON). The linguistic structure of INT and CON are quantitatively investigated on three different but interrelated levels of analysis : Lexis, Syntax and Information Structure. The corpus of data 1n vest1gated for the project consists of eight sessions of pairs of conversants in carefully planned interviews followed by unplanned, surreptitiously recorded conversational encounters of the same pairs of speakers. The data comprise a total of approximately 15.200 words of INT talk and of about 19.200 words in CON. Taking account of the debatable assumption that the language of SES might be complex on certain linguistic levels (e.g. syntax) (Halliday 1979) and might be simple on others (e.g. lexis) in comparison to written discourse, the thesis sets out to investigate this complexity using a statistical approach to the computation of the structures recurrent in the language of INT and CON. The findings indicate clearly the presence of linguistic complexity in both types. They also show the language of INT to be slightly more syntactically and lexically complex than that of CON. Lexical density seems to be relatively high in both types of spoken discourse. The language of INT seems to be more complex than that of CON on the level of information structure too. This is manifested in the greater use of Inferable and other linguistically complex entities of discourse. Halliday's suggestion that the language of SES is syntactically complex is confirmed but not the one that the more casual the conversation is the more syntactically complex it becomes. The results of the analysis point to the general conclusion that the linguistic complexity of types of SES is not only in the high recurrence of syntactic structures, but also in the combination of these features with each other and with other linguistic and extralinguistic features. The linguistic analysis of the language of SES can be useful in understanding and pinpointing the intricacies of spoken discourse in general and will help discourse analysts and applied linguists in exploiting it both for theoretical and pedagogical purposes.

Neuromorphic model for sound source segregation

While humans can easily segregate and track a speaker's voice in a loud noisy environment, most modern speech recognition systems still perform poorly in loud background noise. The computational principles behind auditory source segregation in humans is not yet fully understood. In this dissertation, we develop a computational model for source segregation inspired by auditory processing in the brain. To support the key principles behind the computational model, we conduct a series of electro-encephalography experiments using both simple tone-based stimuli and more natural speech stimulus. Most source segregation algorithms utilize some form of prior information about the target speaker or use more than one simultaneous recording of the noisy speech mixtures. Other methods develop models on the noise characteristics. Source segregation of simultaneous speech mixtures with a single microphone recording and no knowledge of the target speaker is still a challenge. Using the principle of temporal coherence, we develop a novel computational model that exploits the difference in the temporal evolution of features that belong to different sources to perform unsupervised monaural source segregation. While using no prior information about the target speaker, this method can gracefully incorporate knowledge about the target speaker to further enhance the segregation.Through a series of EEG experiments we collect neurological evidence to support the principle behind the model. Aside from its unusual structure and computational innovations, the proposed model provides testable hypotheses of the physiological mechanisms of the remarkable perceptual ability of humans to segregate acoustic sources, and of its psychophysical manifestations in navigating complex sensory environments. Results from EEG experiments provide further insights into the assumptions behind the model and provide motivation for future single unit studies that can provide more direct evidence for the principle of temporal coherence.

Word boundary modelling and full covariance gaussians for Arabic Speech-to-Text systems

This paper describes recent improvements to the Cambridge Arabic Large Vocabulary Continuous Speech Recognition (LVCSR) Speech-to-Text (STT) system. It is shown that wordboundary context markers provide a powerful method to enhance graphemic systems by implicit phonetic information, improving the modelling capability of graphemic systems. In addition, a robust technique for full covariance Gaussian modelling in the Minimum Phone Error (MPE) training framework is introduced. This reduces the full covariance training to a diagonal covariance training problem, thereby solving related robustness problems. The full system results show that the combined use of these and other techniques within a multi-branch combination framework reduces the Word Error Rate (WER) of the complete system by up to 5.9% relative. Copyright © 2011 ISCA.