710513An ultra-efficient decoder for automatic speech recognition (Cantab Research)ClosedGRD Proof of ConceptInnovate UKSpeech recognition and associated services are becoming an increasingly lucrative area of business, and are emerging as increasingly common and crucial components of modern technologies such as smart phones and tablets. Currently popular large vocabulary speech recognition decoders are inefficient in terms of both memory and CPU usage and do not easily fit with the recent success of recurrent neural network language models (RNNLMs). This project will address these issues to create a decoder that will produce results in real time on modest hardware and with vastly reduced memory and processor requirements whilst integrating RNNLMs. This will allow efficient implementation on smartphones and make cloud based speech recognition cheaper and more environmentally friendly whilst maintaining state of the art performance. The project will bring together: * the proposer's technology from a recent Recurrent Neural Net Language Modelling project * the proposer's 25 year experience in this area (inc. authoring of two previous commercial decoders) * four years of design work on this concept