Real time speech / speaker recognition by using digital cochlea system

在本研究中，我们讨论了基于数字耳蜗模型的语音/说话人识别系统。数字耳蜗机模型的优化我们对安装在数字信号处理器上的数字耳蜗机进行了优化。它由行波滤波、速度变换滤波和二次滤波共16个部分组成。数字耳蜗语音/说话人识别算法的研究提出了动态时间规整算法和增强相邻输出之间的差值。它们可以提高识别性能和对噪声的鲁棒性。数字耳蜗片的实现我们在32块DSP板上设计了数字耳蜗板。数字信号处理器TMS320C3xDSK。利用该数字耳蜗式滤波器可以进行实时语音处理。在实时语音识别中的应用我们测试了在各种噪声环境下使用该系统进行的实时语音识别。实验结果表明，在静音环境下，识别率达到99.2%。10dBSXR时为90.6%，5dBSXR时为41.0%。在实时说话人识别中的应用我们使用该系统对18个人的实时说话人识别进行了检验。达到了92.2%的说话人识别率。此外，我们还通过调整每一段数字耳蜗滤波器的增益达到了98.9%。

In this research, we discuss the speech/speaker recognition system using Digital Cochlear Model as follows.1. Optimization of Digital Cochlear ModelWe optimize the Digital Cochlear Model for installation on DSP. It has 16 section consists of traveling-wave filter, velocity transformation filter and second filter.2. Investigation of speech/speaker recognition algorithm for Digital CochleaWe propose the Dynamic Time Warping algorithm and the enhancement of difference between adjacent outputs of Digital Cochlea. They can improve the recognition performance and the robustness against noise.3. Realization of Digital Cochlea filterWe design the digital cochlea on 32 DSP boards. TMS320C3xDSK. Real-time speech processing can be done by this Digital Cochlear filter.4. Application to real-time speech recognitionWe examine the real-time speech recognition using this system under various noisy environment. From experimental results, we achieve 99.2% recognition rate under silent environment. 90.6% under 10dB SXR and 41.0% under 5dB SXR.5. Application to real-time speaker recognitionWe examine the real-time speaker recognition for 18 persons using this system. We achieve 92.2% speaker recognition rate. In addition, we achieve 98.9% by adjusting the gain of each section of Digital Cochlear filter.