[Univ of Cambridge] [Dept of Engineering]


Steve Young










Current and Recent Projects


Open Domain Statistical Spoken Dialogue Systems

The aim of this project is to develop tools and techniques which enable a spoken dialogue system to learn on-line how to sustain a conversation about hitherto unseen concepts and topics. This includes learning to handle new 'slots' being introduced into an existing domain (for example, whether or not a smartphone supports contactless payment) or the emergence of a new domain (for example, internet-connected components for home automation). The techniques being explored are focussed on data-driven components for understanding, dialogue management and generation; and the key ideas centre on the use of generic models which can be specialised to new slots and domains, and committees of experts which can combine knowledge from a pool of domains in order to create a solution for a new domain. ( EPSRC ODS-SDS Project Website)

Statistical Spoken Dialogue Systems for Wide-Domain Applications

This collaborative project in co-operation with Toshiba Cambridge Research Laboratory complements the EPSRC Open Domain SDS project by providing an industrial relevance to the research. Particular focus is on transfer learning from general information domains such as Tourist info to specific product oriented domains such as Laptops and Televisions. The research also includes work on using recurrent neural networks to build trainable NLG.
Sponsor: Toshiba

Robust Dialogue for Infotainment

The goal of this project is to develop techniques and tools which will enable information services to be provided in-car using natural spoken language [ie not keywords or prescribed commands] as the primary input/output modality. The approach is based on a novel statistical framework called the Partially Observable Markov Decision Process (POMDP). The key benefit of the approach is that for a given level of recognition accuracy, a POMDP-based dialogue manager can provide increased tolerance to errors and a more natural user-oriented dialogue. The research results will be demonstrated by a proof-of-concept server-based implementation of an in-car tourist information service which will demonstrate both a habitable dialogue and significantly improved robustness compared to existing spoken dialogue technology.
Sponsor: General Motors

Parlance: Tools for Ambient Linguistic Knowledge

This collaborative project was aimed at designing and building mobile applications that approach human performance in conversational interaction. The project involved developing an architecture to support incremental dialogue, and advancing statistical approaches to allow dialogues to adapt and extend to new domains. ( Parlance Project Website)
Sponsor:EU FP7: STREP