CS224s Next Steps

CS224s Next Steps

  1. Improve our model’s performance

  2. Improve the autoencoder

  3. More training on Autoencoder

  4. Train autoencoder on VCTK

  5. Tune hyperparams

  6. Set up a way to log results

  7. Add Config params as run metadata in tensorboard

  8. Choose most promising params to tune

  9. Improve spectrogram -> waveform conversion at the end

  10. I don’t really understand the process…

  11. Gather examples

  12. Error analysis

  13. Presentation

  14. Paper writeup

  15. Poster

We are trying to optimize for the l2 loss between correlations between filter values on two spectrograms. 

Could we identify formants?