How to update
Replace `recognise_from_audio_files` with `recognise_from_audio_files_with_offsets` in your code or
use the new function `recognise_from_audio_files`, if you don't need the offsets.
Details
Changed the signature of the `recognise_from_audio_files` function.
This function now takes only the path to the audio files and the configuration.
If you need the offsets, use the new function `recognise_from_audio_files_with_offsets`.
Added an ability to "improve" the offsets.
The new step (see `interval_improver`) does two things:
- Filters out the intervals that are too long.
The maximum length of the interval can be set in the configuration (MAX_SEGMENT_LENGTH_SEC).
- Adjusts the start and end of the intervals to the boundaries of the audio files.
This is useful when the found interval are in 2 seconds from the start of the audio file, for example.
The behaviour can be disabled by setting the configuration parameter `ADJUSTMENT_THRESHOLD` to False.
Fixed a bug where the offset was rounded to the nearest integer.
Fixed a warning in the clustering algorithm.
To find the best number of clusters, the algorithm was iterating over the number
of clusters from 2 to the number of offsets. The warning was caused by the
algorithm trying to find more clusters than there are unique offsets.
Now, the algorithm will try to find the best number of clusters from 2 to the
number of unique offsets.
Added a note about the OMP_NUM_THREADS environment variable to the README.
Logging now uses formatted strings instead of f-strings.