Changelog

v0.1.0a

v0.1.0a3

2021-04-09

Bug fixes

• no longer use Librosa for feature extraction: allow easier implementation of augmentations, especially during training.

• soundpy.feats.plot now uses parameter subprocess to allow for different backends to be applied, depending on when funciton is called. For example, if plotting from within a Generator while training, subprocess should be set to True, and the ‘Agg’ backend will be applied. Otherwise, ‘TkAgg’ backend is used. Fixes issues with multi-threading.

• Fixed generator and Tensorflow issue: with Tensorflow 2.2.0+ the models in soundpy.models.builtin that were trained via generator failed. Use tensorflow.data.Dataset.from_generator to feed generator data to models.

• Improved clip_at_zero.

Features

• Python 3.8 can now be used.

• throw depreciation warning for parameters context_window or frames_per_sample as these “features” will be removed from feature extraction. Rather the features can be reshaped post feature extraction.

• added timestep, axis_timestep, context_window, axis_context_window and combine_axes_0_1 paremeters to soundpy.models.Generator: allow more control over shape of the features.

• can run soundpy.models.builtin.envclassifier_extract_train to run with pre-extracted val and test features.

• soundpy.feats.plotsound, soundpy.feats.plot_vad and soundpy.feats.plot_dom_freq all can plot stereo sound: for each channel in a stereo signal, a plot is either generated or saved. If a filename already exists, a date stamp is added to filename to avoid overwriting images.

• allow grayscale2color to be applied to 2D data.

Breaking changes

• soundpy.models.Generator uses parameter normalize instaed of normalized. Found this to be more intuitive. If normalize is set to True, data will be normalized. Before, if normalized was set to True, data would not be normalized.

• removed add_tensor_last and add_tensor_first: require adding of tensors (for keras) to be included in parameter desired_input_shape.

Other changes

• CPU soundpy can use Tensorflow 2.1.0, 2.2.0 and 2.3.0. Dockerfile still uses Tensorflow 2.1.0 as it is still compatible with updated code.

• soundpy.models.builtin.implement_denoiser raises warning if cleaned features cannot be converted to raw audio samples.

v0.1.0a2

2020-08-13

Bug fixes

• added use_beg_ms parameter in soundpy.dsp.vad: improved VAD recognition of silences post speech.

Features

• added GPU option: provide instructions and Docker image for running SoundPy with GPU

• added extend_window_ms paremeter to soundpy.feats.get_vad_samples and soundpy.feats.get_vad_stft: can extend VAD window if desired. Useful in higher SNR environments.

• added soundpy.feats.get_samples_clipped and soundpy.feats.get_stft_clipped to clip off beginning and ending silences.

• added beg_end_clipped parameter to soundpy.feats.plot_vad to visualize VAD by clipping the beginning and ending silences (if True) or VAD instances throughout the signal (if False).

• added soundpy.models.dataprep.GeneratorFeatExtraction class for extracting and augmenting features during training (still experimental).

• added soundpy.models.builtin.envclassifier_extract_train as an example of extracting and augmenting features during training (still experimental).

• added soundpy.dsp.clip_at_zero to enable smoother concatenations of signals and enables removal of clicks at beginning and ending of signals.

• added soundpy.dsp.remove_dc_bias to enable smoother concatenations of signals

• added and set remove_dc parameter to True in soundpy.files.loadsound and soundpy.files.savesound to ensure signals all have mean zero.

• added mirror_sound option to soundpy.dsp.apply_sample_length as a way to extend sound.

• added soundpy.dsp.ismono to check if samples were mono or stereo.

• added soundpy.dsp.average_channels to average sample amplitudes across channels, e.g. to identify where high energy begins / ends in the signal without disregarding additional channels (if stereo sound).

• added soundpy.dsp.add_channels for adding additional channels if needed (e.g. for applying a ‘hann’ or ‘hamming’ window to stereo sound)

• added stereo sound functionality to soundpy.dsp.add_backgroundsound, soundpy.dsp.clip_at_zero, soundpy.dsp.calc_fft, soundpy.feats.get_stft, soundpy.feats.get_vad_stft

Other changes

• name change: from pysoundtool to soundpy: simpler

• updated dependencies to newest versions still compatible with Tensorflow 2.1.0

• moved soundpy.dsp.get_vad_samples to soundpy.feats.get_vad_samples

• moved soundpy.dsp.get_vad_stft to soundpy.feats.get_vad_stft

• name change: allow soundpy.feats.normalize to be used as soundpy.normalize

• removed pysoundtool_online and mybinder button as maintaining the online version was not easily done. Aim to reimplement at some point.

v0.1.0a1

Initial public alpha release.