Exploiting Hidden-Layer Responses of Deep Neural Networks for Language Recognition

The most popular way to apply Deep Neural Network (DNN)for Language IDentification (LID) involves the extraction ofbottleneck features from a network that was trained on automaticspeech recognition task. These features are modeled usinga classical I-vector system. Recently, a more direct DNNapproach was proposed, it consists of estimating the languageposteriors directly from a stacked frames input. The final decisionscore is based on averaging the scores for all the frames fora given speech segment. In this paper, we extended the directDNN approach by modeling all hidden-layer activations ratherthan just averaging the output scores. One super-vector per utteranceis formed by concatenating all hidden-layer responses.The dimensionality of this vector is then reduced using a PrincipalComponent Analysis (PCA). The obtained reduce vectorsummarizes the most discriminative features for languagerecognition based on the trained DNNs. We evaluated this approachin NIST 2015 language recognition evaluation. The performancesachieved by the proposed approach are very competitiveto the classical I-vector baseline.