An Unsupervised frame Selection Technique for Robust Emotion Recognition in Noisy Speech

Automatic emotion recognition with good accuracy has been demonstrated for clean speech, but the performance deteriorates quickly when speech is contaminated with noise. In this paper, we propose a front-end voice activity detector (VAD)-based unsupervised method to select the frames with a relatively better signal to noise ratio (SNR) in the spoken utterances. Then we extract a large number of statistical features from low-level audio descriptors for the purpose of emotion recognition by using state-of-art classifiers. Extensive experimentation on two standard databases contaminated with 5 types of noise (Babble, F-16, Factory, Volvo, and HF-channel) from the Noisex-92 noise database at 5 different SNR levels (0, 5, 10, 15, 20dB) have been carried out. While performing all experiments to classify emotions both at the categorical and the dimensional spaces, the proposed technique outperforms a Recurrent Neural Network (RNN)-based VAD across all 5 types and levels of noises, and for both the databases.