List view

• Productionize classifier

  From final proposal: Check-in with Dr. Lee’s team about current state of their infrastructure Combine the high and low frequency classifiers into a single system that takes an audio sample and correctly classifies all bat calls within

  Overdue by 3 year(s)

  •

  Due by March 15, 2023
  0% complete0 open0 closed

• Distinguish bat calls

  From the proposal: Distinguish types of call using high frequency bat calls Frequency clip audio samples above 30KHz Classify feeding buzz from the rest using unsupervised learning transformers Attempt low hanging fruit search call classification using domain knowledge. Automatically classify any chain of bat calls previous to a buzz feed where each bat call in the chain is within a temporal distance threshold from the others. [Milestone 1] Evaluate with the Bat team (Dr. Lee et al.) if this solution is good for production. If it isn't determined good enough. Develop supervised learning transformers with overlapping sliding windows. [Milestone 2] Distinguish bat call types at all frequencies If the high-frequency classifier generalizes, we’re done (probably not the case, since lower frequency is a lot noisier and calls may not be isomorphic across frequency domain) If the high-frequency classifier does not generalize, repeat steps from 1.a. Output intervals should be in a format consumable by RavenPro [Milestone 3] Implement Test

  Overdue by 3 year(s)

  •

  Due by February 15, 2023

  •1/1 issues closed
  100% complete0 open1 closed

• Data prep and pre-training

  From the proposal: 1. Set up Azure storage and upload .wav files from the ecology research team to S3 bucket. [Dec 19- Dec 31] 2. Identity false positive from annotated data by Raven Pro (number of samples?)[Jan 1 - Jan 14]

  Overdue by 3 year(s)

  •

  Due by January 16, 2023

  •12/13 issues closed
  92% complete1 open12 closed