FeatureExtractor

class audioflux.FeatureExtractor(transforms, num=None, radix2_exp=12, samplate=32000, low_fre=None, high_fre=None, bin_per_octave=12, slide_length=None, scale_type=SpectralFilterBankScaleType.LINEAR, wavelet_type=WaveletContinueType.MORSE)

Batch feature extraction

Parameters

transforms: list or str

Transform type.

Supported: bft/nsgt/cwt/pwt/cqt/st/fst/dwt/wpt

num: int or None

Number of frequency bins to generate.

only BFT/NSGT/CWT/PWT available
cqt is a fixed value of 84
dwt see: DWT default value
wpt see: WPT default value

radix2_exp: int

fft_length=2**radix2_exp

samplate: int

Sampling rate of the incoming audio.

low_fre: float or None

Lowest frequency.

only BFT/NSGT/CWT/PWT available

high_fre: float or None

Highest frequency.

only BFT/NSGT/CWT/PWT available

bin_per_octave: int

Number of bins per octave.

only the Octave type of BFT/NSGT/CWT/PWT and CQT is available

slide_length: int or None

Window sliding length.

only BFT/CQT available

scale_type: SpectralFilterBankScaleType

Spectral filter bank type. It determines the type of spectrogram.

only BFT/NSGT/CWT/PWT available

See: type.SpectralFilterBankScaleType

wavelet_type: WaveletContinueType

Wavelet Type

only CWT available

See: type.WaveletContinueType

Examples

Get a 880Hz’s audio file

>>> import audioflux as af
>>> sample_path = af.utils.sample_path('880')
>>> audio_arr, sr = af.read(sample_path)

Create FeatureExtractor object and extract spectrogram

>>> from audioflux.type import SpectralFilterBankScaleType
>>> fa_obj = af.FeatureExtractor(transforms=['bft', 'cwt', 'cqt'], samplate=sr, radix2_exp=12,
>>>                              scale_type=SpectralFilterBankScaleType.OCTAVE)
>>> spec_result = fa_obj.spectrogram(audio_arr, is_continue=True)

Extract spectral/xxcc/deconv

>>> spectral_result = fa_obj.spectral(spec_result, spectral='flux',
>>>                                   spectral_kw={'is_positive': True})
>>> xxcc_result = fa_obj.xxcc(spec_result, cc_num=13)
>>> deconv_result = fa_obj.deconv(spec_result)

Methods

`deconv`(spec_result[, spec_convert])	Deconv feature
`spectral`(spec_result, spectral[, ...])	spectral related features
`spectrogram`(data_arr[, is_continue])	Get the spectrogram of transforms
`xxcc`(spec_result[, cc_num, rectify_type, ...])	cepstral coefficients

spectrogram(data_arr, is_continue=False)

Get the spectrogram of transforms

Parameters

data_arr: np.ndarray [shape=(n,)] or list(np.ndarray [shape=(n,)])

Input audio datas

is_continue: bool

Calculate continuous 2D(nsgt/cwt/pwt/st/fst/dwt/wpt) spectrogram.

Calculate the cwt every fft/2, then divide the x-axis of each cwt into four parts, and take the middle two parts for splicing (the first part will be spliced with the first part of the first cwt, and the tail will be spliced with the last part of the last cwt).

If True, then calculate continuous 2D spectrogram. If False, the 2D spectrogram can only calculate fft_length(2**radix2_exp) data

Returns

out: FeatureResult: spectrogram data

xxcc(spec_result, cc_num=13, rectify_type=CepstralRectifyType.LOG, spec_convert=<ufunc 'absolute'>)

cepstral coefficients

See: XXCC

Parameters

spec_result: FeatureResult: spectrogram result
cc_num: int: xxcc num, usually set to 13, 20 or 40
rectify_type: CepstralRectifyType: rectify type
spec_convert: function: Operations on spectrogram data

Returns

out: FeatureResult: xxcc data

deconv(spec_result, spec_convert=<ufunc 'absolute'>)

Deconv feature

See: Deconv

Parameters

spec_result: FeatureResult: spectrogram result
spec_convert: function: Operations on spectrogram data

Returns

out: FeatureResult: deconv data

spectral(spec_result, spectral, spectral_kw=None, spec_convert=<ufunc 'absolute'>)

spectral related features

Parameters

spec_result: FeatureResult

spectrogram result

spectral: str

spectral feature name

See: Spectral

spectral_kw: dict or None

spectral feature parameters

See: Spectral

spec_convert: function

Operations on spectrogram data

Returns

out: FeatureResult: spectral data