Parallel Architectures for Speech Recognition: Nonlinear Optimization of Expectation-Maximization (EM) Training of Hidden Markov Models (HMMs) in a Reconfigurable Environment

本研究涉及高速计算新架构的开发、构建和测试。利用的思想是以松耦合的方式，将通用的、基于risc的处理节点与多个基于现场可编程逻辑阵列（FPLA）的协处理器系统相结合。有了这种系统，人们就有了通用计算机的硬件特性，但由于适当的可重构性，其性能更接近于专用引擎。特别是，20个节点的阿姆斯特朗III系统已经建成并投入使用，比目前先进的工作站提供了近两个数量级的改进。硬件系统和它的组态编译器是这类体系结构的基本组成部分。组态编译器是一种自动将C函数子程序转换为机器码和可重构协处理器的硬件设计的程序。因此，这项研究直接集中在国家关注的战略领域之一，即高性能计算。这个项目的独特之处在于，它不仅构建了硬件/软件系统，而且还在重要的实际应用中进行了测试。特别是，一个基于隐马尔可夫模型（HMM）的现代语音识别系统的训练需要数百小时，即使是最近在布朗大学和其他地方开发的一些变体。将训练时间减少到10分钟左右意味着该领域的研究进展可以以更快的速度进行，并且/或者现在可以将昂贵的非线性优化技术应用于该问题。此外，这意味着从一个不那么麻烦的传感器系统（布朗大学正在开发的麦克风阵列系统）获得的数据可以适当地整合到一个更强大的语音识别系统中。目前有8名博士生、2名硕士生和5名本科生参与该项目；在过去的资助下，该项目已培养出6名博士。

This research involves the development, construction and testing of new architectures for high-speed computing. The idea being exploited is that of combining in a loosely-coupled fashion, general-purpose, RISC-based processing nodes each with multiple field-programmable-logic-array(FPLA) based coprocessor systems. With this kind of system, one has the hardware properties of a general-purpose computer, but the advantage, brought about by suitable reconfigurability, of performance more akin to that of a special-purpose engine. In particular, the 20-node Armstrong III system has been built and is operational and provides nearly two-orders of magnitude improvement over current advanced workstations. The hardware system, combined with its configuration compiler, a program that automatically translates a C function subroutine to both machine code and to the hardware design of the reconfigurable coprocessors, are the basic building blocks for this class of architectures. Thus, this research is directly focused on one of the strategic areas of national concern, that of high performance computing. There is uniqueness to this program in that not only has the hardware/software system been built, but it is also being tested on important, real applications. In particular, the training of a modern hidden Markov model(HMM) based speech recognition system requires hundreds of hours, even with some of the variants developed recently at Brown and elsewhere. The reduction of this training time to ten or so minutes means that research progress in this area may be made at a faster rate and/or that expensive nonlinear optimization techniques may now be applied to the problem. Also, this means that data from a less cumbersome sensor system (the microphone array systems in place and being developed at Brown) can be suitably incorporated into a more robust speech recognition system. Currently, eight PhD students, two ScM students and five undergraduates are involved with this project; ni ne PhD's have been graduated from this project under past funding.