I-Vector Representation Based on GMM and DNN for Audio Classification

Explore the state-of-the-art I-vector approach for audio classification tasks in this 57-minute lecture by Najim Dehak from the Center for Language & Speech Processing at Johns Hopkins University. Delve into the modeling and capturing of variability in Gaussian Mixture Model (GMM) mean components across audio recordings. Discover recent subspace approaches like Non-negative Factor Analysis (NFA) and Subspace Multinomial Model (SMM) that focus on GMM weights variability. Learn how these techniques can be applied to model hidden layer neuron activations in deep neural networks for sequential data recognition tasks such as language and dialect recognition. Gain insights from Dehak's extensive background in artificial intelligence, pattern recognition, and speech processing, including his pivotal role in developing the i-vector framework for speaker verification.