Changelog¶
v2.1.0¶
Features¶
Integrate BeatsDROP in SAMPLE with decisors.
Plot beats as dashed lines in plots.
No zero-padding by default.
Nested collapsible HTML menu to inspect model arguments in IPython.
Officially supporting Python 3.11.
Fixes¶
Refactor encapsulation of models to be compliant with the sklearn API.
This also affects argument names. These changes are non-breaking, since old
names are still supported, although deprecated (a warning will be printed).
Adjust amplitude optimization boundaries based on initial values (ensure feasibility
of starting point).
Limit numpy version to avoid issue #1138 in skopt.
Performance¶
Optionally fit BeatsDROP to trajectories in parallel (multiprocessing).
Use overlapping_windows to compute time-frames.
More robust margins in beat regression bounds.
GUI¶
Plot beats as dashed lines in GUI.
Add zero-padding toggle to GUI settings.
Thread lock for analysis.
Major refactoring of GUI internals to accomodate multiprocessing BeatsDROP.
v2.0.0¶
Features¶
Psychoacoustic functions¶
Implementation of conversion functions between Hz and Bark, Mel, ERB, ERB-rate scale (Cams).
Helper functions for decibels.
Cochleagram computation as output of a gammatone filter-bank.
Generalized time-frequency representations plots.
Perceptual losses: Mel-spectrogram Loss, Multiscale Spectral Loss, Cochleagram Loss.
Linear algebra¶
Decorator for functions operating on numpy arrays for automating the allocation of the output array, to avoid multiple allocations.
Strided convolution utility for faster low-resolution convolution outputs.
Optimization¶
Hyperparameter optimization wrapper for Bayesian optimizaiton of SAMPLE parameters.
Beats¶
Random generator of beating modal partials. Modal beat model. DualBeat regression.
GUI¶
Add spacers between parameters sections.
Add “Export WAV” button.
Notebooks¶
Added “Automatic Tuning of SAMPLE hyperparameters” notebook.
v1.5.1¶
Fixes issues with Zenodo metadata
v1.5.0¶
GUI¶
This version introduces a gui intended for a non-developer audience. Also, a stand-alone executable is built for Windows on GitHub
Features¶
Introduced new parameter max_n_modes for limiting the number of modes
used by the SAMPLE model for resynthesis
Fixes¶
Made default boundary computation more robust to degenerate inputs. Also an exception is raised for tracks of length one, advising to raise the minimum sine duration
v1.4.0¶
Features¶
Frequency reduction¶
A custom function can now be provided to reduce the array of frequency of each track into a single frequency. This is mainly intended to provide a way to switch between mean and median
Settable properties¶
Learned properties (frequencies, decays, amplitudes) can now be setted. This allows post-fit manipulations, e.g. pitch-shift
Improvements¶
Now forcing correct package version in notebooks. Also using quiet mode instead of clearing output
v1.3.0¶
Features¶
Merging strategy¶
Introduced an option to decide when to merge two different tracks. The “average” strategy considers the two average frequencies. The “single” strategy considers the last frequency of the first track and the first frequency of the second track
Track strip¶
Introduced an option to strip tracks that start after a given time
v1.2.3¶
Quality-of-code changes:
streamlined workflow
pushed coverage to 100%
added badges to README
v1.2.2¶
Removed most magics from notebooks to turn them crossplatform.
Download is performed via a GET request using requests.
The wav audio format is used for better support.
v1.2.1¶
Minor corrections to docs
v1.2.0¶
Features¶
Reverse mode¶
Added reverse mode: sinusoidal model will (optionally) process audio in reverse. This option was already implemented in the prototypal software for the paper, but missing from this package.
BugFixes¶
HingeRegression default linear_regressor¶
HingeRegression instances with default linear_regressor used to share the same linear_regressor instance. Now assignment is performed on a deepcopy of the parameter value (via property setter).
v1.1.0¶
First official release