Speech Recognition Israel 3 - Prof. Khudanpur & Prof. Keshet

We're excited and honored to host two leading professors in our third meetup:

The first one is Prof. Sanjeev Khudanpur, A world-leading scientist in the area of Speech Recognition that came to a visit in Israel from John Hopkins University.
The second one is Prof. Yossi Keshet From Bar-Ilan University, an expert in Machine Learning and in speech and language processing.

Location: WeWork ToHa, Derech Hashalom 5, 14th Floor
Time: December 29th, 2019. 18:00-20:00

Agenda:
18:00 - 18:30 - Networking (including light food and drinks)
18:30 - 19:10 - Prof. Khudanpur - Two Ideas for Improving Automatic Speech Recognition: One Elegant, and One Very Useful!
19:10 - 19:20 - Break
19:20 - 20:00 - Prof. Keshet - Speech Applications in the Land of Adversity: Attacks, Detection, and Beyond

Abstract for Prof. Khudanpur lecture:
The Kaldi tools for automatic speech recognition (ASR) are being widely used both for research (academic and industry) and for large-scale deployments. I will share vignettes of the inner workings of the Kaldi team by describing how two ideas played out from conceptualization to execution and evaluation. One was to use adversarial examples to improve the training of deep neural networks. The other was GPU acceleration of the inference engine (i.e. Viterbi decoding). They illustrate two different ways in which doing research can be very gratifying: the adversarial training solution turns out to be very elegant, while the GPU acceleration turns out to be of immense practical significance. Time permitting, I will outline optical character recognition capabilities in Kaldi, and other ongoing research threads.

Abstract for Prof. Keshet lecture:
Deep learning has been amongst the most emerging fields in computer science and engineering. In recent years it has been shown that deep networks are vulnerable to attacks by adversarial examples. I will introduce a novel flexible approach named Houdini for generating adversarial examples for complex and structured tasks and demonstrate successful attacks on different applications such as speech recognition, pose estimation, semantic image segmentation, speaker verification, and malware detection.
Then I will discuss how this weakness can be turned into two secure applications. The first is a new technique for watermarking deep network models in a black-box way. That is, concealing information within the model that can be used by the owner of the model to claim ownership. The second application is a novel method to Speech Steganography, namely hiding a secret spoken message within an ordinary public spoken message.
I will conclude the talk by a brief discussion of our attempts to detect such adversarial attacks, based on multiple semantic label representations.