ezmsg.sigproc.filterbank

Functions

filterbank(kernels, mode=FilterbankMode.CONV, min_phase=MinPhaseMode.NONE, axis='time', new_axis='kernel')

Perform multiple (direct or fft) convolutions on a signal using a bank of kernels.

This is intended to be used during online processing, therefore both direct and fft convolutions use the overlap-add method.

Returns: FilterbankTransformer.

Parameters:

• kernels (list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | tuple[ndarray[tuple[Any, ...], dtype[_ScalarT]], ...])

• mode (FilterbankMode)

• min_phase (MinPhaseMode)

• axis (str)

• new_axis (str)

Return type:

FilterbankTransformer

Classes

class Filterbank(*args, settings=None, **kwargs)

Bases: BaseTransformerUnit[FilterbankSettings, AxisArray, AxisArray, FilterbankTransformer]

Parameters:

settings (Settings | None)

SETTINGS

alias of FilterbankSettings

class FilterbankMode(*values)

Bases: OptionsEnum

The mode of operation for the filterbank.

CONV = 'Direct Convolution'

FFT = 'FFT Convolution'

AUTO = 'Automatic'

class FilterbankSettings(kernels: list[numpy.ndarray[tuple[typing.Any, ...], numpy.dtype[~_ScalarT]]] | tuple[numpy.ndarray[tuple[typing.Any, ...], numpy.dtype[~_ScalarT]], ...], mode: ezmsg.sigproc.filterbank.FilterbankMode = <FilterbankMode.CONV: 'Direct Convolution'>, min_phase: ezmsg.sigproc.filterbank.MinPhaseMode = <MinPhaseMode.NONE: 'No kernel modification'>, axis: str = 'time', new_axis: str = 'kernel')

Bases: Settings

Parameters:

• kernels (list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | tuple[ndarray[tuple[Any, ...], dtype[_ScalarT]], ...])

• mode (FilterbankMode)

• min_phase (MinPhaseMode)

• axis (str)

• new_axis (str)

kernels: list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | tuple[ndarray[tuple[Any, ...], dtype[_ScalarT]], ...]

mode: FilterbankMode = 'Direct Convolution'

“conv”, “fft”, or “auto”. If “auto”, the mode is determined by the size of the input data. fft mode is more efficient for long kernels. However, fft mode uses non-overlapping windows and will incur a delay equal to the window length, which is larger than the largest kernel. conv mode is less efficient but will return data for every incoming chunk regardless of how small it is and thus can provide shorter latency updates.

min_phase: MinPhaseMode = 'No kernel modification'

If not None, convert the kernels to minimum-phase equivalents. Valid options are ‘hilbert’, ‘homomorphic’, and ‘homomorphic-full’. Complex filters not supported. See scipy.signal.minimum_phase for details.

axis: str = 'time'

The name of the axis to operate on. This should usually be “time”.

new_axis: str = 'kernel'

The name of the new axis corresponding to the kernel index.

__init__(kernels, mode=FilterbankMode.CONV, min_phase=MinPhaseMode.NONE, axis='time', new_axis='kernel')

Parameters:

• kernels (list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | tuple[ndarray[tuple[Any, ...], dtype[_ScalarT]], ...])

• mode (FilterbankMode)

• min_phase (MinPhaseMode)

• axis (str)

• new_axis (str)

Return type:

None

class FilterbankState

Bases: object

tail: ndarray[tuple[Any, ...], dtype[_ScalarT]] | None = None

template: AxisArray | None = None

dest_arr: ndarray[tuple[Any, ...], dtype[_ScalarT]] | None = None

prep_kerns: ndarray[tuple[Any, ...], dtype[_ScalarT]] | list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | None = None

windower: WindowTransformer | None = None

fft: Callable | None = None

ifft: Callable | None = None

nfft: int | None = None

infft: int | None = None

overlap: int | None = None

mode: FilterbankMode | None = None

class FilterbankTransformer(*args, **kwargs)

Bases: BaseStatefulTransformer[FilterbankSettings, AxisArray, AxisArray, FilterbankState]

class MinPhaseMode(*values)

Bases: OptionsEnum

The mode of operation for the filterbank.

NONE = 'No kernel modification'

HILBERT = 'Hilbert Method; designed to be used with equiripple filters (e.g., from remez) with unity or zero gain regions'

HOMOMORPHIC = 'Works best with filters with an odd number of taps, and the resulting minimum phase filter will have a magnitude response that approximates the square root of the original filter’s magnitude response using half the number of taps'

class FilterbankMode(*values)

Bases: OptionsEnum

The mode of operation for the filterbank.

CONV = 'Direct Convolution'

FFT = 'FFT Convolution'

AUTO = 'Automatic'

class MinPhaseMode(*values)

Bases: OptionsEnum

The mode of operation for the filterbank.

NONE = 'No kernel modification'

HILBERT = 'Hilbert Method; designed to be used with equiripple filters (e.g., from remez) with unity or zero gain regions'

HOMOMORPHIC = 'Works best with filters with an odd number of taps, and the resulting minimum phase filter will have a magnitude response that approximates the square root of the original filter’s magnitude response using half the number of taps'

class FilterbankSettings(kernels: list[numpy.ndarray[tuple[typing.Any, ...], numpy.dtype[~_ScalarT]]] | tuple[numpy.ndarray[tuple[typing.Any, ...], numpy.dtype[~_ScalarT]], ...], mode: ezmsg.sigproc.filterbank.FilterbankMode = <FilterbankMode.CONV: 'Direct Convolution'>, min_phase: ezmsg.sigproc.filterbank.MinPhaseMode = <MinPhaseMode.NONE: 'No kernel modification'>, axis: str = 'time', new_axis: str = 'kernel')

Bases: Settings

Parameters:

• kernels (list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | tuple[ndarray[tuple[Any, ...], dtype[_ScalarT]], ...])

• mode (FilterbankMode)

• min_phase (MinPhaseMode)

• axis (str)

• new_axis (str)

kernels: list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | tuple[ndarray[tuple[Any, ...], dtype[_ScalarT]], ...]

mode: FilterbankMode = 'Direct Convolution'

“conv”, “fft”, or “auto”. If “auto”, the mode is determined by the size of the input data. fft mode is more efficient for long kernels. However, fft mode uses non-overlapping windows and will incur a delay equal to the window length, which is larger than the largest kernel. conv mode is less efficient but will return data for every incoming chunk regardless of how small it is and thus can provide shorter latency updates.

min_phase: MinPhaseMode = 'No kernel modification'

If not None, convert the kernels to minimum-phase equivalents. Valid options are ‘hilbert’, ‘homomorphic’, and ‘homomorphic-full’. Complex filters not supported. See scipy.signal.minimum_phase for details.

axis: str = 'time'

The name of the axis to operate on. This should usually be “time”.

new_axis: str = 'kernel'

The name of the new axis corresponding to the kernel index.

__init__(kernels, mode=FilterbankMode.CONV, min_phase=MinPhaseMode.NONE, axis='time', new_axis='kernel')

Parameters:

• kernels (list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | tuple[ndarray[tuple[Any, ...], dtype[_ScalarT]], ...])

• mode (FilterbankMode)

• min_phase (MinPhaseMode)

• axis (str)

• new_axis (str)

Return type:

None

class FilterbankState

Bases: object

tail: ndarray[tuple[Any, ...], dtype[_ScalarT]] | None = None

template: AxisArray | None = None

dest_arr: ndarray[tuple[Any, ...], dtype[_ScalarT]] | None = None

prep_kerns: ndarray[tuple[Any, ...], dtype[_ScalarT]] | list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | None = None

windower: WindowTransformer | None = None

fft: Callable | None = None

ifft: Callable | None = None

nfft: int | None = None

infft: int | None = None

overlap: int | None = None

mode: FilterbankMode | None = None

class FilterbankTransformer(*args, **kwargs)

Bases: BaseStatefulTransformer[FilterbankSettings, AxisArray, AxisArray, FilterbankState]

class Filterbank(*args, settings=None, **kwargs)

Bases: BaseTransformerUnit[FilterbankSettings, AxisArray, AxisArray, FilterbankTransformer]

Parameters:

settings (Settings | None)

SETTINGS

alias of FilterbankSettings

filterbank(kernels, mode=FilterbankMode.CONV, min_phase=MinPhaseMode.NONE, axis='time', new_axis='kernel')

Perform multiple (direct or fft) convolutions on a signal using a bank of kernels.

This is intended to be used during online processing, therefore both direct and fft convolutions use the overlap-add method.

Returns: FilterbankTransformer.

Parameters:

• kernels (list[ndarray[tuple[Any, ...], dtype[_ScalarT]]] | tuple[ndarray[tuple[Any, ...], dtype[_ScalarT]], ...])

• mode (FilterbankMode)

• min_phase (MinPhaseMode)

• axis (str)

• new_axis (str)

Return type:

FilterbankTransformer