Robust Front-End Processing For Emotion Recognition In Noisy Speech
Meghna Abhishek Pandharipande, Rupayan Chakraborty, Ashish Kumar Panda, Sunil Kumar Kopparapu · 2018
Since the emotion recognition performances degrade drastically for noisy speech, we propose a robust front-end processing to reduce the effect of noise. The front-end consists of a novel energy based voice activity detector (VAD), which discards silence or noisy frames. We show that the proposed VAD results in significant performance gain and outperforms the more complex Recurrent Neural Network (RNN) based VAD as well as popular Non-negative Matrix Factorization (NMF) technique. Moreover, the proposed VAD can be used alongside the NMF technique to further improve the performance. The emotion recognition is done by extracting a large number of statistical features from low-level audio descriptors, and we have used state-of-the-art classifiers. Extensive experimentation on noisy (5 types of noise: Babble, F-16, Factory, Volvo, and HF-channel from the Noisex-92 database) speech contaminated at 5 different SNR levels (0,5,10,15,20dB) have been carried out to measure the performance of the proposed front-end techniques.