to-do-list.txt

CS224s TO-DO list:

1. Figure out a way to distinguish between genuine and bogus predictions (i.e. for defined and undefined pitch values respectively).

2. Optimize training parameters for network with all features by plotting network performance as a function of manipulating these parameters.

3. Decide if to remove final output layer conflating BW_o and FW_o. Why not feed the BW_o and FW_o directly into the regression & regularization? This would recast the task as a multiple regression problem and makre regularization more appropriate. 

4. Extract output weights from network before they are fed to linear regression and run 10-fold cross-validation to select best Lasso regularization lambda. Actually since the regression only has 1 variable at the moment (or 2 if we remove final output layer), then ridge is more justified, because with so few variables we don't want to remove any of them. We just want to optimize the weights.

5. Train network after removing a single feature family at a time to see how important a family of features is in boosting network performance (generate 5%, 10%, 25% true mean SD classifications). 

6. Create a more even train-dev-test split (e.g. 80-10-10) to respect standards in the field.

7. Transplant pitch track from human speech samples to kal's synthesized voice to obtain more directly comparable true labels.

8. Generate Festival utterances for testing, extract F0s and compare how well it does compared to true labels (generate 5%, 10%, 25% true mean SD classifications) to compare with our results.

9. Interpolate and smooth predicted F0 (possibly using cubic spline interpolation; find out if periodic parameter of spline function in R more appropriate given pitch data).

10. Remove parts of pitch track between undefined F0 points (can probably keep single undefined points sandwiched between two defined points)

10. Generate sound files from synthesized continuous pitch tracks. Align with syllable and utterance durations. Insert syllable and utterance labels (i.e. text) into the appropriate tiers.
 
11. Analyse more Festival synthesized files in terms of AER & BER to get more robust accuracy statistics. Sample with replacement to improve estimates.

11. Calculate how many hours of recordings your dataset has. Break that into training-dev-test sets as well. This is a more standard way of reporting dataset size in spoken language processing research.

12. Calculate AER & BER for our network's predicted F0 after interpolation and smoothing.

13. In your results table caption, note that the prediction accuracy ignored undefined F0 values.