Reza Yazdani Aminabadi

PhD Researcher


Welcome to my website

I am a PhD researcher at Computer Architecture department of Universitat Politecnica de Catalunya. I study my PhD under the supervision of Jose-Maria Arnau and Antonio Gonzalez at ARCO (ARchitecture and COmpiler) group. My current research is focusing on designing energy-efficient and high-performance hardware accelerators for Automatic Speech Recongnition (ASR) systems.

I started my PhD at 2015. I specifically work on different stages of ASR pipeline which are mainly the acoustic likelihood computation (DNN) and speech graph search (Viterbi Search). As the first work, I build an accelerator for the Viterbi search, as the main bottleneck of ASR decoders, which ouperforms both GPU and CPU significantly and reduces the system's energy by several orders of magnitude. Secondly, we focused on the training phase of generating the speech graphs. By separating the large dataset of large-vocabulary ASR graphs together with some novel graph compression techniques, we could reduce the memory footprint from more than 1 GB to less than 40 MB! You can find out a summary of my work on ASR systems here. New!


News

  • Confidence in DNNs: Our new paper of evaluating and handling this new metric has been accepted in ISCA'18.
  • DNN Pruning, Adoption toward Adaption!
    Evaluating the Negative side-effects of DNN pruning in different AI application.
  • Exploring DNN and RNN for the hybrid speech recognition systems