#include <spectral.h>

Detailed Description

template<typename Sample>
class signalsmith::spectral::STFT< Sample >

STFT synthesis, built on a MultiBuffer.

Any window length and block interval is supported, but the FFT size may be rounded up to a faster size (by zero-padding). It uses a heuristically-optimal Kaiser window modified for perfect-reconstruction.

Simulated bad-case aliasing (random phase-shift for each band) for overlapping ratios

There is a "latest valid index", and you can read the output up to one historyLength behind this (see .resize()). You can read up to one window-length ahead to get partially-summed future output.

You move the valid index along using .ensureValid(), passing in a functor which provides spectra (using .analyse() and/or direct modification through .spectrum[c]):

void processSample(...) {
    stft.ensureValid([&](int) {
        // Here, we introduce (1 - windowSize) of latency
        stft.analyse(inputBuffer.view(1 - windowSize))
    });
    // read as a MultiBuffer
    auto result = stft.at(0);
    ++stft; // also moves the latest valid index
}
 
void processBlock(...) {
    // assuming `historyLength` == blockSize
    stft.ensureValid(blockSize, [&](int blockStartIndex) {
        int inputStart = blockStartIndex + (1 - windowSize);
        stft.analyse(inputBuffer.view(inputStart));
    });
    auto earliestValid = stft.at(0);
    auto latestValid = stft.at(blockSize);
    stft += blockSize;
}

The index passed to this functor will be greater than the previous valid index, and <= the index you pass in. Therefore, if you call .ensureValid() every sample, it can only ever be 0.

Inherits signalsmith::delay::MultiBuffer< Sample >.

Inherited by signalsmith::spectral::ProcessSTFT< Sample >.

Types
enum class	Window { kaiser , acg }

using	Spectrum = MultiSpectrum

Types inherited from signalsmith::delay::MultiBuffer< Sample >
using	ConstChannel = typename Buffer< Sample >::ConstView

using	MutableChannel = typename Buffer< Sample >::MutableView

using	MutableView = View< false >

using	ConstView = View< true >

Static Attributes
static constexpr Window	kaiser = Window::kaiser

static constexpr Window	acg = Window::acg

Methods
void	setWindow (Window shape, bool rotateToZero=false)
	Swaps between the default (Kaiser) shape and Approximate Confined Gaussian (ACG). More...

	STFT (int channels, int windowSize, int interval, int historyLength=0, int zeroPadding=0)
	Parameters passed straight to `.resize()` More...

void	resize (int nChannels, int windowSize, int interval, int historyLength=0, int zeroPadding=0)
	Sets the channel-count, FFT size and interval. More...

int	windowSize () const

int	fftSize () const

int	interval () const

decltype(fft.window())	window () const
	Returns the (analysis and synthesis) window. More...

std::vector< Sample >	partialSumWindow (bool includeLatestBlock=true) const
	Calculates the effective window for the partially-summed future output (relative to the most recent block) More...

void	reset ()
	Resets everything - since we clear the output sum, it will take `windowSize` samples to get proper output. More...

template<class AnalysisFn >
void	ensureValid (int i, AnalysisFn fn)
	Generates valid output up to the specified index (or 0), using the callback as many times as needed. More...

template<class AnalysisFn >
void	ensureValid (AnalysisFn fn)
	The same as above, assuming index 0. More...

int	nextInvalid () const
	Returns the next invalid index (a.k.a. the index of the next block) More...

template<class Data >
void	analyse (Data &&data)
	Analyse a multi-channel input, for any type where `data[channel][index]` returns samples. More...

template<class Data >
void	analyse (int c, Data &&data)

template<class Data >
void	analyseRaw (Data &&data)
	Analyse without windowing or zero-rotation. More...

template<class Data >
void	analyseRaw (int c, Data &&data)

int	bands () const

int	latency ()
	Internal latency (between the block-index requested in `.ensureValid()` and its position in the output) More...

STFT &	operator++ ()

STFT &	operator+= (int i)

STFT &	operator-- ()

STFT &	operator-= (int i)

Super::MutableView	operator++ (int postIncrement)

Super::MutableView	operator-- (int postIncrement)

Methods inherited from signalsmith::delay::MultiBuffer< Sample >
	MultiBuffer (int channels=0, int capacity=0)

void	resize (int nChannels, int capacity, Sample value=Sample())

void	reset (Sample value=Sample())

Stride< false >	at (int offset)

Stride< true >	at (int offset) const

MutableView	view (int offset=0)

ConstView	view (int offset=0) const

ConstView	constView (int offset=0) const

MutableChannel	operator[] (int channel)

ConstChannel	operator[] (int channel) const

MultiBuffer &	operator++ ()

MultiBuffer &	operator+= (int i)

MutableView	operator++ (int)

MutableView	operator+ (int i)

ConstView	operator+ (int i) const

MultiBuffer &	operator-- ()

MultiBuffer &	operator-= (int i)

MutableView	operator-- (int)

MutableView	operator- (int i)

ConstView	operator- (int i) const

Attributes
Window	windowShape = Window::kaiser

Spectrum	spectrum

WindowedFFT< Sample >	fft

Constructor & Destructor Documentation

◆ STFT()

template<typename Sample >

signalsmith::spectral::STFT< Sample >::STFT	(	int	channels,
		int	windowSize,
		int	interval,
		int	historyLength = `0`,
		int	zeroPadding = `0`
	)

inline

Parameters passed straight to .resize()

Method Details

◆ analyse()

template<typename Sample >

template<class Data >

void signalsmith::spectral::STFT< Sample >::analyse ( Data && data )

inline

Analyse a multi-channel input, for any type where data[channel][index] returns samples.

Results can be read/edited using .spectrum.

◆ analyseRaw()

template<typename Sample >

template<class Data >

void signalsmith::spectral::STFT< Sample >::analyseRaw ( Data && data )

inline

Analyse without windowing or zero-rotation.

◆ ensureValid() [1/2]

template<typename Sample >

template<class AnalysisFn >

void signalsmith::spectral::STFT< Sample >::ensureValid ( AnalysisFn fn )

inline

The same as above, assuming index 0.

◆ ensureValid() [2/2]

template<typename Sample >

template<class AnalysisFn >

void signalsmith::spectral::STFT< Sample >::ensureValid	(	int	i,
		AnalysisFn	fn
	)

inline

Generates valid output up to the specified index (or 0), using the callback as many times as needed.

The callback should be a functor accepting a single integer argument, which is the index for which a spectrum is required.

The block created from these spectra will start at this index in the output, plus .latency().

◆ latency()

template<typename Sample >

int signalsmith::spectral::STFT< Sample >::latency ( )

inline

Internal latency (between the block-index requested in .ensureValid() and its position in the output)

Currently unused, but it's in here to allow for a future implementation which spreads the FFT calculations out across each interval.

◆ nextInvalid()

template<typename Sample >

int signalsmith::spectral::STFT< Sample >::nextInvalid ( ) const

inline

Returns the next invalid index (a.k.a. the index of the next block)

◆ partialSumWindow()

template<typename Sample >

std::vector<Sample> signalsmith::spectral::STFT< Sample >::partialSumWindow ( bool includeLatestBlock = true ) const

inline

Calculates the effective window for the partially-summed future output (relative to the most recent block)

◆ reset()

template<typename Sample >

void signalsmith::spectral::STFT< Sample >::reset ( )

inline

Resets everything - since we clear the output sum, it will take windowSize samples to get proper output.

◆ resize()

template<typename Sample >

void signalsmith::spectral::STFT< Sample >::resize	(	int	nChannels,
		int	windowSize,
		int	interval,
		int	historyLength = `0`,
		int	zeroPadding = `0`
	)

inline

Sets the channel-count, FFT size and interval.

◆ setWindow()

template<typename Sample >

void signalsmith::spectral::STFT< Sample >::setWindow	(	Window	shape,
		bool	rotateToZero = `false`
	)

inline

Swaps between the default (Kaiser) shape and Approximate Confined Gaussian (ACG).

Default (Kaiser) windows and partial cumulative sum

The ACG has better rolloff since its edges go to 0:

ACG windows and partial cumulative sum

However, it generally has worse performance in terms of total sidelobe energy, affecting worst-case aliasing levels for (most) higher overlap ratios:

Simulated bad-case aliasing for ACG windows - compare with above

Roughly optimal Kaiser for STFT analysis (forced to perfect reconstruction)

◆ window()

template<typename Sample >

decltype(fft.window()) signalsmith::spectral::STFT< Sample >::window ( ) const

inline

Returns the (analysis and synthesis) window.

Member Data Documentation

◆ windowShape

template<typename Sample >

Window signalsmith::spectral::STFT< Sample >::windowShape = Window::kaiser

Deprecated:: use .setWindow() which actually updates the window when you change it

Detailed Description

template<typename Sample> class signalsmith::spectral::STFT< Sample >

Types

Static Attributes

Methods

Attributes

Constructor & Destructor Documentation

◆ STFT()

Method Details

◆ analyse()

◆ analyseRaw()

◆ ensureValid() [1/2]

◆ ensureValid() [2/2]

◆ latency()

◆ nextInvalid()

◆ partialSumWindow()

◆ reset()

◆ resize()

◆ setWindow()

◆ window()

Member Data Documentation

◆ windowShape

template<typename Sample>
class signalsmith::spectral::STFT< Sample >