Massively parallel computing in natural language technology

虽然过去20年来语音和语言处理最著名的趋势可能仍然是采用统计方法来破解各种以前难以处理的计算应用程序(语音识别、机器翻译、网页搜索等)，但该领域最近转向了另一个非常不同、不太知名但同样具有变革性的角落。对这些应用中许多应用的最新研究现在与其说是通过它们的全新技术，不如说是因为它们的根本更大的**规模**，当然，创新仍然在不断涌现。很多机器翻译团队基本上仍在使用IBM上世纪90年代早期的模型，但训练了超过0.5 PB(兆兆字节)的数据。事实上，最好的文本到语音合成器处理语音信号的方式与15年前的前辈大致相同，但它们将数据库中记录的语音片段粘贴在一起，这些数据库比当时的数据库大了数万倍。现在做得更大，实际上对生产更好的产品有很大帮助。

While the best known trend in speech and language processing over the last twenty years is probably still its adoption of statistical methods to crack open a variety of previously intractable computing applications (speech recognition, machine translation, web-page search, etc.), this field has recently turn another, very different, less celebrated, but equally transformative corner. State of the art research on many of these applications can now be spotted not so much by their radically new techniques, although innovations have of course still been forthcoming, as by their radically larger **scales**. A great many machine translation groups still essentially use IBM's models from the early 1990s, but trained over half a petabyte (a megabyte of megabytes) of data. The best text-to-speech synthesizers manipulate speech signals in about the same way as their predecessors did 15 years ago, in fact, but they paste together pieces of recorded speech from databases that are tens of thousands of times larger than they were back then. Bigger now does actually go a very long way towards producing better.