All code under the klio/integration directory (including this job) exists for the purpose of integration testing changes to the klio ecosystem and will not work as a regular deployed job.
If you are looking for an example of a working klio job, please look in the klio/examples directory.
This test adapts this public librosa example for separating vocals from instrumental into a klio pipeline.
- Use v2 of configuration (and defacto v2 of klio messages)
- Read klio messages from a local text file
- Using the new transforms in
klio-audio:- Download audio into memory from GCS
- Load audio from memory into librosa
- Calculate the short-time fourier transform (STFT)
- Separate vocals from instruments (multiple "user"-implemented transforms)
- Using new klio transforms in
klio-audio:- Generate matplotlib figures (png) for the separation
- Upload the plots to GCS
- Runs on direct runner
- The output file matches the input file
- No de/serialization issues
- For each entity ID, there are three resulting plots:
- the full magnitude spectrogram
- the background spectrogram (instrumental)
- the foreground spectrogram (vocal)
In the top level of this repo:
$ KLIO_TEST_DIR=audio-spectrograms tox -c integration/tox.ini