Feature compensation based on the normalization of vocal tract length for the improvement of emotion-affected speech recognition

Abstract The performance of speech recognition systems trained with neutral utterances degrades significantly when these are tested emotional speech. Since everybody can speak emotionally in the real-world environment, it is necessary to take account states automatic system. Limited works have been performed field emotion-affected and so far, most researches focused on classification emotions. In this paper, vocal tract length normalization method employed enhance robustness For purpose, two structures system based hybrids hidden Markov model Gaussian mixture deep neural network used. To achieve goal, frequency warping applied filterbank and/or discrete-cosine transform domain(s) feature extraction process conducted a way normalize components make them close their corresponding components. proposed evaluated neutrally trained/emotionally conditions for different features (i.e., Anger , Disgust Fear Happy Sad ). system, acoustical features. constructed Kaldi Persian database crowd-sourced multi-modal actors dataset as input corpora. experimental simulations reveal that, general, warped result better compared unwarped counterparts. Also, be seen that using network-hidden outperforms employing hybrid model.