adithyabsk / bird_audio Goto Github PK

Pre-Processing

• High-pass filtering

Processing / Extracting features

• Check that librosa accepts mp3

• Extract features based on the gender paper

• Read about ts-fresh

• time series --> static features

Modelling

• Logistic regression
• Loss (chart)
  • validation
  • train
  • Make sure to get the final values stored somewhere
• Look into feature importance / PCA (nice have)

• Visualization
  • Song vs Call over time (of year) Heat map
  • Map of audio on the US
  • Histogram of english name
    • stack by song vs call
  • nans per column
    • nice to have

• uninstall kaggle
• write script to scrape from xenocanto (https://www.xeno-canto.org/explore?query=type%3Asoundscape) or cornell (https://search.macaulaylibrary.org/catalog?tag=env&mediaType=a)
  • is there enough density of location data over time (different times of day)? (so we can answer the question: is changing bird song in a location over time an indicator of pollution)
  • pull bird call metadata
  • pull sonograms
  • pull audio files
• data integration with DVC
• write notes on data pipelining using DVC
• pull weather data from OpenWeather API based on recording locations

Blocked by #10

• Start final report doc in deepnote
• Further reading on bioacoustics as markers of environmental stressors, other research using soundscape and/or weather data (see the Future work sections)
• Read up on quality control - how data gets into xentocanto / Cornell dataset
• Come up with more research questions that involve soundscape and/or weather data, what qualities we'd need to normalize in the data for each
• Background on specific questions (gender, location, clustering types of birdcalls)

• metadata
• sono

• Build dataset featurization script (after exploration in jupyter)
• Model to answer: Can we classify calls vs songs in bird audio?

(400 words)

• Motivate the research (why do we care about birdcalls)
• What does xenocanto data look like (what data are we actually using) - "data dictionary" - key features of the data (xenocanto API) e.g. location, type of call
• Xentocanto data quality

All the models need to be done

• Is this a good question to answer?
• Can we do it ourselves
• Update the project document