一种用于说话人辨认的概率神经网络的MCE训练算法

MCE-based PNN Training Algorithm for Speaker Identification

提出了一种基于最小分类错误准则的概率神经网络的训练算法。实验结果表明 ,这种分类网络及其学习算法在 30个说话人辨认应用中利用 5秒清晰语音获得 98.4%的辨认率 ,利用 15秒电话语音获得 85 .1%的辨认率。

This paper presents a training algorithm for probabilistic neural networks using the MCE criterion. Experimental results for speaker identification indicate that this classifier network and its training algorithm attain 98.4% identification accuracy using 5 second clean speech utterances and 85.1% accuracy using 15 second telephone speech utterances with a 10 speaker population.