Investigate Data Preprocessing Alternatives For Spectrogram/CNN Approach

[KatJHuang]

For our CNN model to return a prediction, we must supply the input in the form of an image. But our raw input is an audio file. To which party (Google Cloud, Android, etc.) to delegate this data format conversion is a problem to be solved. We should investigate alternatives to achieve this utilitarian functionality.

Possible Alternatives:

• Processing on Android: https://bitbucket.org/galmiza/spectrogram-android Looks like ppl have developed apps to compute spectrograms. We can give this a try.
• Google Dataflow: https://cloud.google.com/ml-engine/docs/tensorflow/flowers-tutorial here is an example of how to do data transformation with GCP. Might be an overkill and costs a bit.
• Build a server (E.g. Heroku) that just does audio to image conversion. Python script is ready for use in this case. Just need to deploy it somewhere and somehow connect with Google Cloud Function and Google ML Engine. Now that I think about it this is actually pretty feasible as workload is less, i.e. not porting existing tools to other platforms. The caveat is that the audio packages needed for my python scripts may not be supported on Heroku servers.