Abstract:  The Kaldi tools for automatic speech recognition (ASR) are being widely used both for research and for industry-scale deployments.  Many innovations keep Kaldi up-to-date in this fact-moving field.  This presentation will briefly overview the history of ASR and Kaldi, then describe how two recent innovations grew from germination to implementation and evaluation.  One is to use adversarial examples to improve training of deep neural network (DNN) based acoustic models.  The other is GPU acceleration of the inference engine (the so-called Viterbi decoder).  They illustrate two different metrics of research success.  The adversarial training solution turns out to be very elegant: it can be viewed as either correcting a biased estimate of the true gradient in SGD, or as a form of classical leave-one-out estimation.  The GPU acceleration solution turns out to be of immense practical significance, and demonstrates a successful academic-industry collaboration in applied areas of Computer Science.

Bio: https://www.clsp.jhu.edu/faculty-pages/sanjeev