FeaturesBase Class

Abstract Base Class for audio feature extraction. Defines an interface for audio feature extraction algorithms.

Inheriting classes must override get_features() and implement feature extraction in that function, which accepts an AudioBuffer and must return a np.ndarray containing results.

Feature extraction is run in inheriting classes through the __call__ member function.

Example of FFT, which inherits from FeaturesBase:

audio = AudioBuffer('./some_audio.wav')
fft = FFT()

# Now feature extraction is run by treating
# the FFT instance, fft, like a function
spectrum = fft(audio)
class spiegelib.features.FeaturesBase(sample_rate=44100, time_major=False, scale=False, scale_axis=(0, ))

Bases: abc.ABC

Parameters

  • sample_rate (int, optional) – audio sample rate, defaults to 44100

  • time_major (bool, optional) – for audio feature extraction that uses time slices, used to indicate the orientation of features and time slices in the resulting matrix. True indicates that results should be returned with an orientation like (time_slices, features). False (default) refers to an orientation of (features, time_slices).

  • scale (bool, optional) – whether to scale the results of feature extraction. The scaler must be set before scaling called be applied. See fit_scaler() or set_scaler().

  • scale_axis (int, tuple, None, optional) – indicates the axis to use for fitting scalers and applying data scaling. A value of None flattens the dataset and calculates scaling parameters on that. A value of an int or a tuple indicates the axis or axes to use. Defaults to (0,) which causes scaling variables to be calculated on each feature independently.

__call__(audio, scale=None)

Run this feature extraction pipeline.

Applies functions in this order: input modifiers > feature extraction > prescale modifiers > data scaling > output modifiers

Parameters

  • audio (AudioBuffer) – Audio to extract features from

  • scale (bool, optional) – If set, will override scale attribute set during construction.

Returns

results from audio feature extraction with modifiers and scaling.

Return type

np.ndarray

abstract get_features(audio)

Must be implemented. Run audio feature extraction on audio provided as parameter. Must check the time_major attribute and return data in the correct orientation.

Parameters

audio (AudioBuffer) – Audio to extract features from

Returns

Results of audio feature extraction

Return type

np.ndarray

add_modifier(modifier, type)

Add a data modifier to the feature extraction pipeline.

Input modifiers are applied to raw AudioBuffers <audio_buffer> prior to feature extraction. Prescale modifiers are applied to results of audio feature extraction and before data scaling (if applicable). Output modifiers are applied after data scaling.

Parameters

  • modifier (lambda) – data modifier function. Should accept an np.array, apply some modification to that, and return a np.array

  • type (str) – Where to add function into pipeline. Must be one of (‘input’, ‘prescale’, or ‘output’)

fit_scaler(data, transform=True)

Fit scaler to dataset for future transforms.

Parameters

  • data (np.ndarray) – data to train (fit) scaler on

  • transform (bool, optional) – if True will also apply scaling to the data

Returns

Scaled data if transform parameter is True, otherwise returns None.

Return type

np.ndarray, None

has_scaler()
Returns

whether or not the scaler has been set

Return type

bool

set_scaler(scaler)

Set a scaler for a dimension, this will be used to normalize that dimension

Parameters

scaler (object) – a Scaler object to use to scale data. Must be a DataScalerBase type to use the fit_scaler() method. Otherwise, other scalers like sklearn scalers could potentially be fit and then passed in here – just needs to have a transform method to apply scaling. (note, using other objects outside of DataScalerBase type has not been tested!)

scale(data)

Scale features using pre-trained scaler

Parameters

data (np.ndarray) – data to be scaled

Returns

scaled data

Return type

np.ndarray

load_scaler(location)

Load trained scaler from a pickled file.

Parameters

location (str) – Location of pickled scaler object

save_scaler(location)

Save the trained scaler for these features as a pickle for later use

Parameters

location (str) – Location to save pickled scaler object