A Study on Robust Speech Recognition System Using Characteristic Space Filtering and its Low-Power LSI Design

本研究开发了一种新的低功耗噪声鲁棒语音识别算法及其实时集成电路系统。这些技术在实际环境中得到了应用，对移动人机界面的开发具有实用价值。在算法开发方面，提出了通用的鲁棒语音识别系统。使用该方法，在0 dB信噪比以上的非特定男女环境中，最佳性能的识别正确率约为98%，平均性能的识别正确率为60%。在大规模集成电路设计方面，设计并实现了用于鲁棒自动语音识别（ASR）的超并行、超流水线处理大规模集成电路。LSI的工艺为0.25微米，其功耗比基于DSP的相同ASR系统可降低1/5。除此之外，还设计了包括该LSI在内的整个系统，即完整的ASR系统板。它具有麦克风IO，模拟放大器和数字ASR系统。在实际应用中可以获得较高的鲁棒性语音识别性能。

In this study, the new algorithm of noise robust speech recognition and its real-time LSI system with low po wer consumption have been developed. These techniques are used under actual environment and useful for mobile human interface. As to the algorithm development, general purpose robust speech recognition system has been proposed. Using this method, about 98% recognition correctness at best performance and 60% correctness at average performance can be obtained for unspecific male and female over 0 dB SNR environments.As to LSI design, super-parallel and super-pipeline processing LSI for robust automatic speech recognition (ASR) has been designed and realized. The technology of LSI is 0.25micron and its power consumption can be reduced 1/5 than DSP based same ASR system. In addition to this design, the total system, i.e., the full ASR system board including this LSI has also been designed. It has microphone IO, analog amplifier and digital ASR system. The high performance of robust speech recognition can be obtained in the real fields.