Realistic speech interface

Computers remain very limited regarding interpretation of natural human communication. One reason is that the main method for instructing computers, expert annotation, is itself slow, expensive and inflexible.

The EU-funded HELENLP (Heterogeneous learning for natural language processing) project aimed to develop systems better able to interact with humans. The project focused on improving performance by adding a vast new range of digital sources, and in varying degrees of annotation. Using the planned algorithms, computers would be able to automatically interpret the sources by querying an annotator. The team intended to develop additional annotation algorithms with the ultimate purpose of improving machine understanding of natural speech and text.

Work focused on new computational and statistical methods for integrating and analysing digital sources. Researchers developed passive ways of annotating partially annotated data, without an annotator, plus active methods using an annotator. The new algorithms also addressed simultaneous learning through the combination of several sources of annotated data. The applications include text categorisation and phoneme segmentation and recognition.

The undertaking achieved its goal of reintegrating a strong research group. Such work yielded a pool of researchers, additional funding and research publications.

HELENLP's new algorithms may help to improve computer processing of natural human communication.