Towards improved performance of learning algorithms on irregular signals
Deep learning has shown outstanding performance in various tasks. However, recent efforts found that introducing a network with a signal with irregularities can cause the network to underperform. In this talk, I will present three works that try to alleviate the performance of neural networks on different irregularities. The first work handles irregularities from adversarial examples. We propose to create a neighborhood (a ball) around each training example, such that the label is kept constant for all inputs within that neighborhood. I will show that this can be expressed as a new loss function that is attack independent. The second work deals with “creaky voice” or “vocal fry”, an irregularity in the voice that became very common among young native speakers of English, and which causes some automatic speech recognition systems to fail. I will present an algorithm to detect and localize the portions of creaky voice in fluent speech. I will conclude the talk with our work on irregularities in the pronunciation of non-native speakers. This work inspects a pair of speech samples spoken by different talkers. We represent the samples using an embedding space and then project them onto a low-dimensional space, in which they create a trajectory. Using the trajectory, we analyze and measure the divergence between speakers for two cases: (i) accentedness and intelligibility, and (ii) accentedness and accent errors in code-switching.
* Roni Chernyak is a Ph.D. candidate at the Electrical Engineering Department at the Technion under the supervision of Prof. Yossi Keshet.