Analog CMOS Neural Circuits for Speech Recognition

与说话人无关的连接语音的实时识别一直是电子系统的一个难题。低精度高并行模拟神经网络已被证明可以识别使用标准组件的依赖于说话人的小词汇连接语音。该项目将研究和开发大规模并行定制集成电路，使用集成延迟线和模拟计算网络来执行这项任务。第一阶段包括测试芯片的设计和制造，以证明我们开发的可行性；第二阶段的努力将通过设计、制造和测试全尺寸原型系统来继续这一发展。该方法利用了标准的CMOS批量集成电路技术，因此由本研究产生的产品可以由许多供应商可靠而经济地制造。第一阶段的高潮是制造和测试一个可扩展的、可编程的单芯片语音分类器。该芯片的工作速度超过100 MIPS，性能比DSP芯片提高了50倍（归一化到相同的硅面积使用）。pi还显示了使用浮动节点提供模拟电压的非易失性存储的可行性。这些电压可用于控制芯片的操作，从而实现在现场学习或训练的能力。在第二阶段，pi将研究学习算法到高效硅电路的映射。本研究拟应用P.I.在实时语音识别这一具有挑战性和重要的问题上，他的经验可能会直接导致具有广泛商业适用性的廉价产品。

Real-time recognition of speaker-independent connected speech has proven to be a difficult problem for electronic systems. Low-precision highly parallel analog neural networks have been shown to recognize speaker-dependent small-vocabulary connected speech using standard components. This project will research and develop massively parallel custom integrated circuits that will perform this task, using integrated delay lines and analog computing networks. Phase I has included the design and fabrication of test chips to prove the feasibility of our development; the Phase II effort will continue this development through design, fabrication and testing of full- scale prototype systems. The approach utilizes standard readily available CMOS Bulk integrated circuit technology so products arising from this research can be fabricated reliably and economically by a number of vendors. Phase I culminated in the fabrication and testing of an expandable, programmable single-chip speech classifier. This chip operates at over 100 MIPS and enjoys a 50x improvement in performance over DSP chips (normalized to the same silicon area usage). The P.I. has also shown the feasibility of using floating-nodes to provide nonvolatile storage of analog voltages. These voltages can be used to control the operation of the chip thus implementing the capability to learn or be trained in the field. During Phase II, the P.I. will investigate the mapping of learning algorithms to efficient silicon circuits. This proposed research applies the P.I.'s experience to the challenging and important problem of real-time speech recognition and may lead directly to inexpensive products with wide commercial applicability.