MRI: Acquisition of Tesla Hardware for Speech Recognition

提案编号：CNS 09-23511 PI(s)：Janin, Adam 机构：国际计算机科学研究所 标题：MRI/Acq.：收购用于语音识别项目的 Tesla 硬件 提议：该项目收购了 nVidia Tesla 架构系统，有助于开发自动语音识别 (ASR) 的并行代码算法。用于语音识别的 Tesla 硬件由 10 个 nVidia Tesla 机架安装单元以及相关基础设施组成。由于计算机系统现在遵循“核心定律”，因此必须开发新的并行代码以提高语音识别的准确性。 （芯片上的核心数量每两年翻一番）。 该系统提供了大规模多核通用计算环境，支持开发可扩展的并行代码算法。该仪器提供了一个测试平台，可用于研究未来的可扩展算法，这些算法预计将有助于在先进和新颖的算法方面促进商业和工业市场的持续领先地位。因此，展望 5 到 10 年后多核架构盛行时的计算状态，该项目的目标是首先确保适当的计算资源用于语音识别研究，然后超越当前的 2 核和 4 核基本桌面系统。该研究所在具有挑战性的声学特性和自然的人与人交流的现实环境中进行机器学习和语音识别的广泛培训。应用程序运行从为残疾用户提供的免提访问和为不懂计算机的人提供的自然语音驱动界面到自动会议助理和浏览器，其中实时记录会议，并提供允许在会议期间和会议后访问内容的工具。以下两种相关方法可以提高准确性： - 多流方法，涉及系统内多个级别的组合，包括多个特征、多个机器学习估计器和多个词流组合； - 增加训练集的大小。即使训练数据与实际应用条件不完全匹配，仔细的数据集成也可以提高准确性。这两种方法都需要提高当前传统硬件难以满足的计算能力。更广泛的影响：此次收购有助于继续吸引年轻研究人员并帮助他们进行培训。计算能力的提高有助于向当地高中生展示语音研究。 BFOIT 信息技术机会基金会旨在吸引更多女性和代表性不足的少数群体进入计算机科学和工程领域。

Proposal #: CNS 09-23511 PI(s): Janin, Adam Institution: International Computer Science InstituteTitle: MRI/Acq.: Acquisition of Tesla Hardware for Speech Recognition Project Proposed:This project, acquiring an nVidia Tesla architecture system, facilitates development parallel code algorithms for automatic speech recognition (ASR). The Tesla Hardware for Speech Recognition consists of a cluster of 10 nVidia Tesla rack-mounted units with associated infrastructure. New parallel codes must be developed to improve the accuracy of speech recognition, since computer systems now obey ?Core?s Law? (where the number of cores on a chip doubles once every two years). The system provides a large-scale multi-core general purpose computing environment that enables the development of scalable parallel code algorithms. The instrument provides a testbed in which to investigate future scalable algorithms that are expected to facilitate continued leadership in the commercial and industrial markets in terms of advanced and novel algorithms. Thus, envisioning the state of computing in 5 to 10 years when multi-core architectures prevail, the project aims to first- Secure appropriate computing resources for research in speech recognition and then- Eclipse the current 2- and 4-core basic desktop systems.The institute performs extensive training for machine learning and speech recognition in realistic settings with challenging acoustic properties and natural, human to human communication. Applications run from hands-free access to disabled users and natural speech-driven interfaces for the non-computer literate to automatic meeting assistants and browsers, in which meetings are recorded in real-time and tools are provided that allow access to content both during and after the meeting. The following two relevant methods improve accuracy:- Multi-stream methods that involve combinations at many levels within the system, including multiple features, multiple machine learning estimators, and multiple word-streams combinations and- Increasing the size of the training set.Careful integration of data can improve the accuracy even when the training data does not exactly match the conditions of actual application. Both methods require increasing computational power hard to fulfill with current conventional hardware.Broader Impacts: The acquisition contributes to continue attracting young researchers and aiding in their training. The improved computational capability facilitates the demonstration of speech research to local high school students. The BFOIT Foundation for Opportunities in Information Technology aims to attract more women and underrepresented minorities in computer science and engineering.