Deep neural networks for anger detection from real life speech data
Jun Deng, Florian Eyben, Björn Wolfgang Schuller, Felix Burkhardt · 2017
There has been a lot of previous work on deep neural networks for automatic speech recognition, however, little emphasis has been placed on an investigation of effective deep learning architectures for anger detection from speech. In this paper, inspired by the state-of-the-art deep learning algorithms, we propose a variant of Deep Long Short-Term Memory (LSTM) Recurrent Neural Networks (RNNs), Convolution Neural Networks (CNNs) with 3 × 3 kernels, and LSTM RNNs combined with CNNs, in conjunction with log-mel filter bank features and brute forced low-level-descriptors from the standardised ComParE set for speech anger detection. We extensively evaluate the deep networks on a big real-life speech corpus of 26 970 utterances with utterance-level labels collected from a German voice portal, finding that our proposed neural networks significantly outperform traditional modelling algorithms for speech anger detection.