Output details
13 - Electrical and Electronic Engineering, Metallurgy and Materials
Queen's University Belfast
Optimization of Weighted Finite State Transducer for Speech Recognition
The paper describes a number of key algorithmic optimisations for an embedded speech recognition system. Traditional speech recognition systems build speech sources on-the-fly, using them to perform recognition involving heavy computation and large bandwidth; this work takes a radically different approach by pre-creating the network to reduce memory and power. The work resulted in UK patent 1020771.0 and was subsequently funded under an Invest Northern Ireland proof-of-concept grant, PoC132. The work won the 2010 HiTech CONNECT £25K Award which aims to identify knowledge-based ideas with the most commercial potential, www.nisp.co.uk/25k-award/.