Large Scale Neural Network Acoustic Models (Cantab Research)

语音建模正在成为一个越来越有利可图的商业领域，尤其是在语音到文本、语音索引、语音验证、说话人识别和文本到语音等领域，所有这些都变得越来越普遍，成为智能手机和平板电脑等现代技术的关键组成部分。目前的性能还远远不够完美，这个项目将创造新的算法来提高这些技术的性能。该项目将汇集：*统计建模的最新进展*用图形处理器单元构建多兆次浮点规模计算机的易用性*机器学习在声学建模中的历史小规模使用*提案人在该领域25年的经验和* Cantab开发的基础研究，尚未在商业规模上进行实验验证。Cantab研究的主要结果将显著提高自动语音识别和许可收入其他领域。

Modelling of speech is becoming an increasingly lucrative area of business, especially in theareas of speech-to-text, speech indexing, speech verification, speaker identification and text-tospeech,all of which are becoming increasingly common and crucial components of moderntechnologies such as smart phones and tablets.Current performance is far from perfect, and this project will create new algorithms toimprove performance of these technologies.The project will bring together:* recent advances in statistical modelling* the ease of building multi-teraflop scale computers with Graphics Processor Units* historical small scale use of machine learning in acoustic modelling* the proposer's 25 year experience in this area and* basic research developed by Cantab that has yet to be experimentally verified on acommercial scaleThe primary result for Cantab Research will be significantly better automatic speechrecognition andlicensing revenue for other domains.