基于VQ-SVM的说话人识别系统

Speaker recognition based on VQ-SVM

说话人识别系统通常是将在安静的环境下训练得到的参数应用于实际环境中。如果实际环境也是安静的,则说话人识别系统可以令人满意地工作。然而,当实际环境中有噪声存在时,说话人识别系统性能急剧下降。为了让说话人识别系统在安静的环境和有噪声的环境中都获得令人满意的工作性能,研究了一个将支持向量机(SVM)在矢量量化(VQ)系统上进行二次识别来提高说话人识别率的方法。通过引入阈值自适应,从而提高系统性能。实验表明,在噪声环境下,与VQ,SVM识别方法相比,此方法在对识别速度影响很小的情况下可以使识别率明显提高。此方法具有良好的应用前景和进一步研究的价值。

Speaker recognition systems worked in practical environments using parameters train in quiet environments. The system performance was satisfactory when the environment was also quiet, but degraded quickly in noisy environments A two - level - architecture recognition method combining VQ and SVM was propce, ed so that the recognition system could work well both in quiet and noisy environments. A threshold was adopted to improve the performance and of the system. The result of the experiment shows that compared with the VQ and SVM recognition method the proposed method could highly improve the recognition rate with little impact on the running speed. It has a good application prospect and worth to research further more.