to-dos

[Rahul-Brito]

• [x] Look at Hausdorff analysis for just the one embedding model
• [x] downsample mother audio
• [x] run new audio through embedding extraction
• [x] Do t-test between non-loo and loo distribution
• [x] Get egemaps features from my data
• [x] Learn CCA to see what im supposed to get
• [x] Expand CCA
• [x] update pyannote to 2.0 to see if it improves results

[Rahul-Brito]

• [x] Listen to samples to make sure it is clean
• [x] See if baby shows up as a second speaker
• [x] See if lab tech shows up as second speaker
• [x] Remove lab tech (just cut 1 min perhaps)
• [x] Next draft of blog. Organize what i need and want in teh context of a paper. "please start preparing the outline of the results on a paper on voice proximity. what is it that you would like the story to be"

[Rahul-Brito]

To do any of the remaining tech items, i need to

• [x] figure out how to remove lab tech from the recording using the pyannote APIs. could just cut out one minute but this seems unreliable, especially if she sometimes talks at the end. I don't want to lose too much data too. Once I do this I can try the above things

• [x] Chunk out gemaps based on diarized audio

• [x] Run he gemaps data through the same pipeline of dim reduc

• [ ] CCA - look at whether F0 is represented in the embeddings or how much

• [ ] Look at distances WITHIN participants - talk to Gasser

• [x] On hausdorff distances - compare each cell (pairwise) instead of distributions

• [x] Sweep different parameters for KNN, etc. (e.g. instead of 10 neighbors, sweep to 50), etc.

• [x] Need to do a better job contextualizing the problem and space

• [ ] Look at silhouette coefficient or rand-initi (?)

[Rahul-Brito]

• [x] look at this question to Herve and his reply. https://github.com/pyannote/pyannote-audio/discussions/923
• [x] look at this one too: https://github.com/pyannote/pyannote-audio/discussions/924

[Rahul-Brito]

• [x] fix metric learning hausdorff code if it needs it
• [x] consider if i should compare all-UMAP to metric learned or version or not? probably should
• [x] finish reading tsne documentation - do I need to scale things first? or is it auto done. Do i need to run PCA first or is it autodone?
• [x] finish high-dim cos analysis

[Rahul-Brito]

• [x] fit on all 20 then fit transform on 19? vizualize these to compare

[Rahul-Brito]

RA feature requests

Yep, that makes sense! The diarization pipeline itself we have works quite well it seems, so we should be able to separate mother from RA voice automatically. I am hoping/assuming we can do this for the baby cries too. I am not sure if any of our recordings got extended baby crying or not to test this, but I can check!

[Rahul-Brito]

• [ ] pull the one emb mod script into the one collab. Pick one dimension reduction - worth comparing UMAP vs TSNE

[Rahul-Brito]

• [ ] do librasa for f0 etc
• [ ] Do something something high dimensions
• [ ] Figure out what the dsitance metrics mean