具有环境自学习机制的鲁棒说话人识别算法

Robust Speaker-Recognition Algorithm with Environmental Self-Learning Mechanism

说话人识别系统实际应用时,一旦应用环境和训练环境不一致,系统的性能会急剧下降。由于环境噪声的多变性,系统训练时无法预测实际应用中的环境噪声。因此,引入环境自学习和自适应思想,通过改进的矢量泰勒级数(Vector Taylor Series,VTS)刻画环境噪声模型和说话人语音模型之间的统计关系,提出一种具有环境自学习能力的鲁棒说话人识别算法。系统应用中每当环境变化时利用语音输入前采集到的环境噪声信号来迭代更新环境噪声模型参数,进一步基于VTS确立的统计关系,将说话人语音模型自适应到实际应用环境来补偿环境失配的影响。说话人辨认实验结果表明,提出的方法在低信噪比条件下对于不同种类的噪声都能显著提升系统的识别性能。

In the actual application of the speaker recognition system,once application environment and the training environment are inconsistent,the performance of the system will drop significantly.Due to the variability of environmental noise,the environmental noise in practical applications cannot be predicted during system training.Therefore,the environment self-learning and adaptive ideas are introduced to describe the statistical relationship between the environmental noise model and the speaker’s speech model through the improved VTS(Vector Taylor Series),and a robust speaker-recognition algorithm with environmental self-learning ability is proposed.In system application,when environment changes,the environment noise before speech input is collected to iteratively update the model parameters of environment noise,and further adapt the speaker model to the application environment to compensate for the environmental mismatch based on the statistical relationship established by VTS.The speaker-recognition experiment results indicate that the proposed method can significantly improve the recognition performance of the system for different kinds of noise under low SNR conditions.