Data Driven Approaches to Speech and Language Processing

Speech and language processing systems can be categorised according to whether they make use of predefined linguistic information and rules or are data driven and therefore exploit machine learning techniques to automatically extract and process relevant units of information which are then indexed and retrieved as appropriate. As an example, most state of the art automatic speech processing systems rely on a representation based on predefined phonetic symbols. The use of language dependent representations, whilst linguistically intuitive, has several drawbacks i.e. portability across languages, development time. Therefore, in this article, we review and present our recent experiments exploiting the idea inherent in the ALISP (Automatic Language Independent Speech Processing) approach, with particular respect to speech processing, where the intermediate representation between the acoustic and linguistic levels area is automatically inferred from speech data. We then present prospective directions in which the ALISP principles could be exploited by different domains such as audio, speech, text, image and video processing.