Parallel Architectures for Speech Recognition: Testing Expensive Algorithms in a Reconfigurable Environment

西尔弗曼 一个强大的研究计划，在先进的语音识别，算法和实验的计算架构领域目前在布朗的工程人/机系统（LEMS）实验室。本研究的重点是一个可重构的，通用的，计算结构和语音识别/训练算法，以推进国家的最先进的在这两个领域通过它们的相互作用。 这些努力的一个特点是评估语音识别器早期阶段变化的影响。 这项计算密集型任务将利用计划中的Armstrong III可重构硬件系统来加速当前的训练算法。 第二个特征是对说话者独立的、连接的字母数字识别器、数据库、训练和工具的当前计算和识别性能的改进的结合，所有这些都是在Brown开发的。这些包括麦克风阵列数据采集，新的信号处理算法，非参数建模，更好的持续建模，和新的VQ/特征空间方法。

Silverman A strong research program in the areas of computing architectures for advanced speech recognition, algorithms, and experimentation is currently in place in the Laboratory for Engineering Man/machine Systems (LEMS) at Brown. This research focuses on a reconfigurable, general-purpose, computing structure, and speech recognition/training algorithms, to advance the state-of-the-art in both areas through their interaction. One feature of these efforts is the evaluation of the effects of changes to the early stages of a speech recognizer. This computationally intensive task will utilize the projected Armstrong III reconfigurable hardware system to accelerate the current training algorithm. A second feature is the incorporation of improvements to the current computational and recognition performance of a talker- independent, connected alphadigit recognizer, database, training, and tools, all developed at Brown. These include microphone-array data acquisition, new signal-processing algorithms, non- parametric modeling, better durational modeling, and new VQ/feature-space methods.