An Analysis of Speech Emotion Recognition Based on Hybrid DNN-HMM Framework

This paper presents the results of using toolkit Kaldi and library Vosk for training and online recognition of isolated words in Serbian language. Kaldi allows the training of acoustic models, while Vosk is a Python library for speech recognition that can efficiently utilize those models. The focus of this research is analysis of robustness of automatic speech recognition (ASR) system trained using neutral speech, and tested against speech under different emotional states. The description of training models in Kaldi from scratch is given, along with parameter fine-tuning and the recognition results. Comparison was made against recognizers based on Hidden Markov models (HMM) and Convolutional Neural Networks (CNN). A noticeable increase in recognition accuracy is obtained for all examined emotional states.