CMUsphinx is a fast and flexible open source speech recognition toolkit. Currently it uses GMM acoustic models and it gives reasonable accuracy. This project is geared towards incorporating a state of the art MLP-HMM model into sphinx, substantially decreasing the error rates. A ResNet with MLPs instead of convolution networks will be trained and tested on TEDLIUM corpus.The aim is to achieve best-in-category accuracy in recognition of Ted talks.This project will bring CMUsphinx at par with other powerful speech toolkits like KALDI and will help Sphinx users worldwide get better accuracy in their ASR projects.



Hammad Abdullah


  • Bhiksha Raj
  • Nikolay Shmyrev