基于区分性准则的Bottleneck特征及其在LVCSR中的应用

Discriminative Criterion Based Bottleneck Feature and Its Application in LVCSR

基于深层神经网络中间层的Bottleneck(BN)特征由于可以采用传统的混合高斯模型-隐马尔可夫建模(Gaussian mixture model-hidden Markov model,GMM-HMM),在大规模连续语音识别中获得了广泛的应用。为了提取区分性的BN特征,本文提出在使用传统的BN特征训练好GMM-HMM模型之后,利用最小音素错误率(Minimum phone error,MPE)准则来优化BN网络参数以及GMM-HMM模型参数。该算法相对于其他区分性训练算法而言,采用的是全部数据作为一个大的数据包,而不是小的包方式来训练深度神经网络,从而可以大大加快训练速度。实验结果表明,优化后的BN特征提取网络比传统方法能获得9%的相对词错误率下降。

Bottleneck （BN） features based on the middle layer of deep neural network have been widly ap‐plicated to large vocabulary continuous speech recognition （LVCSR） ,because they can use the traditional Gaussian mixture density hidden Markov model （GMM‐HMM） for acoustic modeling .In order to extract discriminative bottleneck features ,the parameters of the BN feature extractor and GMM‐HMM are opti‐mized jointly by using the minimum phone error （MPE） criterion after training the GMM‐HMM using the conventional BN features .Different from other discriminative training method ,large batches instead of mini‐batch in conventional neural network optimization are used to obtain the statistics ,which acceler‐ates training speed .Experiments demonstrate that the proposed bottleneck feature extractor can outper‐form the traditional methods with 9% relative word error reduction .