Machine Learning Applied to Speech Technology and Autonomous Agents
UEF Summer School, 2018

Learning diary

Returnables: Filled learning diary (Deadline 9.9)
Filled learning diary will be returned to Moodle (https://moodle.uef.fi/course/view.php?id=17032).

For the first week of the course, participants are required to submit filled learning diary that consists of different questions related to given lectures.

Learning diary questions: https://moodle.uef.fi/mod/resource/view.php?id=1309620

Project (done during second week, 20.8 - 24.8, and onward)

Returnables: Final report and source code (Deadline 9.9)
Report and source code will be returned to the Moodle.

Challenge 1: Train an agent that plays Toribash

  • Toribash is a 1v1 fighting game.
  • Design and/or train an agent for Toribash-DestroyUke-v1 task: The higher cumulative reward at the end of the game/episode, the better. Find the environment from Github link below.
  • Ready made learning environment (ToriLLE) can be found from this link: https://github.com/Miffyli/ToriLLE
  • Information about ToriLLE can be found from our paper https://arxiv.org/abs/1807.10110

Challenge 2: Imitating a celebrity voice

  • Use the corpus VoxCeleb (http://www.robots.ox.ac.uk/~vgg/data/voxceleb/)
  • It contains a very large number (> 1000) of celebrity voices. Find closest to your voice using speaker recognition technology.
  • Try to see if you can get better score by modifying your voice.
  • Get the project package including pre-trained speaker recognition models from here.

Challenge 3: Test the impact of environmental noise in a Speaker Recognition System (SRS)

  • Modify a set of speech signals from Voxceleb by adding noise and test the performance of the SRS (Use the SRS from the practice)
  • Apply a speech enhancement method and recheck the system performance (e.g. Wiener filter in Scipy)
  • Challenge the system with a different issue and compare results (e.g. short duration)
  • Record your own voice and repeat the process. Compare the obtained results with the one with the previous set of signals and discuss.
  • Get the project package including pre-trained speaker recognition models from here.