react-native-audio-api 0.5.6 → 0.6.0-rc.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/RNAudioAPI.podspec +1 -1
- package/android/src/main/cpp/audioapi/android/core/AudioPlayer.cpp +0 -20
- package/android/src/main/cpp/audioapi/android/core/AudioPlayer.h +0 -2
- package/android/src/main/java/com/swmansion/audioapi/AudioAPIPackage.kt +13 -0
- package/android/src/main/java/com/swmansion/audioapi/AudioManagerModule.kt +59 -0
- package/android/src/oldarch/NativeAudioManagerModuleSpec.java +99 -0
- package/common/cpp/audioapi/AudioAPIModuleInstaller.h +2 -6
- package/common/cpp/audioapi/core/AudioContext.cpp +1 -12
- package/common/cpp/audioapi/core/AudioContext.h +0 -1
- package/common/cpp/audioapi/core/sources/AudioBufferSourceNode.cpp +1 -1
- package/common/cpp/audioapi/libs/signalsmith-stretch/fft-accelerate.h +326 -0
- package/common/cpp/audioapi/libs/signalsmith-stretch/fft.h +1257 -413
- package/common/cpp/audioapi/libs/signalsmith-stretch/signalsmith-stretch.h +398 -232
- package/common/cpp/audioapi/libs/signalsmith-stretch/stft.h +625 -0
- package/ios/audioapi/ios/AudioAPIModule.mm +2 -3
- package/ios/audioapi/ios/AudioManagerModule.h +18 -0
- package/ios/audioapi/ios/AudioManagerModule.mm +92 -0
- package/ios/audioapi/ios/core/AudioPlayer.h +4 -12
- package/ios/audioapi/ios/core/AudioPlayer.m +26 -108
- package/ios/audioapi/ios/core/IOSAudioPlayer.h +1 -3
- package/ios/audioapi/ios/core/IOSAudioPlayer.mm +4 -28
- package/ios/audioapi/ios/system/AudioEngine.h +23 -0
- package/ios/audioapi/ios/system/AudioEngine.mm +137 -0
- package/ios/audioapi/ios/system/AudioSessionManager.h +22 -0
- package/ios/audioapi/ios/system/AudioSessionManager.mm +183 -0
- package/ios/audioapi/ios/system/LockScreenManager.h +23 -0
- package/ios/audioapi/ios/system/LockScreenManager.mm +299 -0
- package/ios/audioapi/ios/system/NotificationManager.h +16 -0
- package/ios/audioapi/ios/system/NotificationManager.mm +151 -0
- package/lib/module/api.js +1 -0
- package/lib/module/api.js.map +1 -1
- package/lib/module/core/AudioContext.js +2 -1
- package/lib/module/core/AudioContext.js.map +1 -1
- package/lib/module/specs/NativeAudioManagerModule.js +31 -0
- package/lib/module/specs/NativeAudioManagerModule.js.map +1 -0
- package/lib/module/specs/index.js +6 -0
- package/lib/module/specs/index.js.map +1 -0
- package/lib/module/system/AudioManager.js +66 -0
- package/lib/module/system/AudioManager.js.map +1 -0
- package/lib/module/system/index.js +4 -0
- package/lib/module/system/index.js.map +1 -0
- package/lib/module/system/types.js +2 -0
- package/lib/module/system/types.js.map +1 -0
- package/lib/typescript/api.d.ts +1 -0
- package/lib/typescript/api.d.ts.map +1 -1
- package/lib/typescript/core/AudioContext.d.ts.map +1 -1
- package/lib/typescript/specs/NativeAudioManagerModule.d.ts +13 -0
- package/lib/typescript/specs/NativeAudioManagerModule.d.ts.map +1 -0
- package/lib/typescript/specs/index.d.ts +4 -0
- package/lib/typescript/specs/index.d.ts.map +1 -0
- package/lib/typescript/system/AudioManager.d.ts +12 -0
- package/lib/typescript/system/AudioManager.d.ts.map +1 -0
- package/lib/typescript/system/index.d.ts +2 -0
- package/lib/typescript/system/index.d.ts.map +1 -0
- package/lib/typescript/system/types.d.ts +28 -0
- package/lib/typescript/system/types.d.ts.map +1 -0
- package/package.json +2 -2
- package/src/api.ts +1 -0
- package/src/core/AudioContext.ts +6 -1
- package/src/specs/NativeAudioManagerModule.ts +51 -0
- package/src/specs/index.ts +6 -0
- package/src/system/AudioManager.ts +122 -0
- package/src/system/index.ts +1 -0
- package/src/system/types.ts +68 -0
- package/common/cpp/audioapi/libs/signalsmith-stretch/delay.h +0 -715
- package/common/cpp/audioapi/libs/signalsmith-stretch/perf.h +0 -82
- package/common/cpp/audioapi/libs/signalsmith-stretch/spectral.h +0 -493
|
@@ -1,82 +0,0 @@
|
|
|
1
|
-
#ifndef SIGNALSMITH_DSP_PERF_H
|
|
2
|
-
#define SIGNALSMITH_DSP_PERF_H
|
|
3
|
-
|
|
4
|
-
#include <complex>
|
|
5
|
-
|
|
6
|
-
#if defined(__SSE__) || defined(_M_X64)
|
|
7
|
-
# include <xmmintrin.h>
|
|
8
|
-
#else
|
|
9
|
-
# include <cstdint> // for uintptr_t
|
|
10
|
-
#endif
|
|
11
|
-
|
|
12
|
-
namespace signalsmith {
|
|
13
|
-
namespace perf {
|
|
14
|
-
/** @defgroup Performance Performance helpers
|
|
15
|
-
@brief Nothing serious, just some `#defines` and helpers
|
|
16
|
-
|
|
17
|
-
@{
|
|
18
|
-
@file
|
|
19
|
-
*/
|
|
20
|
-
|
|
21
|
-
/// *Really* insist that a function/method is inlined (mostly for performance in DEBUG builds)
|
|
22
|
-
#ifndef SIGNALSMITH_INLINE
|
|
23
|
-
#ifdef __GNUC__
|
|
24
|
-
#define SIGNALSMITH_INLINE __attribute__((always_inline)) inline
|
|
25
|
-
#elif defined(__MSVC__)
|
|
26
|
-
#define SIGNALSMITH_INLINE __forceinline inline
|
|
27
|
-
#else
|
|
28
|
-
#define SIGNALSMITH_INLINE inline
|
|
29
|
-
#endif
|
|
30
|
-
#endif
|
|
31
|
-
|
|
32
|
-
/** @brief Complex-multiplication (with optional conjugate second-arg), without handling NaN/Infinity
|
|
33
|
-
The `std::complex` multiplication has edge-cases around NaNs which slow things down and prevent auto-vectorisation. Flags like `-ffast-math` sort this out anyway, but this helps with Debug builds.
|
|
34
|
-
*/
|
|
35
|
-
template <bool conjugateSecond=false, typename V>
|
|
36
|
-
SIGNALSMITH_INLINE static std::complex<V> mul(const std::complex<V> &a, const std::complex<V> &b) {
|
|
37
|
-
return conjugateSecond ? std::complex<V>{
|
|
38
|
-
b.real()*a.real() + b.imag()*a.imag(),
|
|
39
|
-
b.real()*a.imag() - b.imag()*a.real()
|
|
40
|
-
} : std::complex<V>{
|
|
41
|
-
a.real()*b.real() - a.imag()*b.imag(),
|
|
42
|
-
a.real()*b.imag() + a.imag()*b.real()
|
|
43
|
-
};
|
|
44
|
-
}
|
|
45
|
-
|
|
46
|
-
#if defined(__SSE__) || defined(_M_X64)
|
|
47
|
-
class StopDenormals {
|
|
48
|
-
unsigned int controlStatusRegister;
|
|
49
|
-
public:
|
|
50
|
-
StopDenormals() : controlStatusRegister(_mm_getcsr()) {
|
|
51
|
-
_mm_setcsr(controlStatusRegister|0x8040); // Flush-to-Zero and Denormals-Are-Zero
|
|
52
|
-
}
|
|
53
|
-
~StopDenormals() {
|
|
54
|
-
_mm_setcsr(controlStatusRegister);
|
|
55
|
-
}
|
|
56
|
-
};
|
|
57
|
-
#elif (defined (__ARM_NEON) || defined (__ARM_NEON__))
|
|
58
|
-
class StopDenormals {
|
|
59
|
-
uintptr_t status;
|
|
60
|
-
public:
|
|
61
|
-
StopDenormals() {
|
|
62
|
-
uintptr_t asmStatus;
|
|
63
|
-
asm volatile("mrs %0, fpcr" : "=r"(asmStatus));
|
|
64
|
-
status = asmStatus = asmStatus|0x01000000U; // Flush to Zero
|
|
65
|
-
asm volatile("msr fpcr, %0" : : "ri"(asmStatus));
|
|
66
|
-
}
|
|
67
|
-
~StopDenormals() {
|
|
68
|
-
uintptr_t asmStatus = status;
|
|
69
|
-
asm volatile("msr fpcr, %0" : : "ri"(asmStatus));
|
|
70
|
-
}
|
|
71
|
-
};
|
|
72
|
-
#else
|
|
73
|
-
# if __cplusplus >= 202302L
|
|
74
|
-
# warning "The `StopDenormals` class doesn't do anything for this architecture"
|
|
75
|
-
# endif
|
|
76
|
-
class StopDenormals {}; // FIXME: add for other architectures
|
|
77
|
-
#endif
|
|
78
|
-
|
|
79
|
-
/** @} */
|
|
80
|
-
}} // signalsmith::perf::
|
|
81
|
-
|
|
82
|
-
#endif // include guard
|
|
@@ -1,493 +0,0 @@
|
|
|
1
|
-
#include <audioapi/core/Constants.h>
|
|
2
|
-
|
|
3
|
-
#ifndef SIGNALSMITH_DSP_SPECTRAL_H
|
|
4
|
-
#define SIGNALSMITH_DSP_SPECTRAL_H
|
|
5
|
-
|
|
6
|
-
#include <audioapi/libs/signalsmith-stretch/perf.h>
|
|
7
|
-
#include <audioapi/libs/signalsmith-stretch/fft.h>
|
|
8
|
-
#include <audioapi/libs/signalsmith-stretch/delay.h>
|
|
9
|
-
|
|
10
|
-
#include <audioapi/dsp/Windows.h>
|
|
11
|
-
|
|
12
|
-
#include <cmath>
|
|
13
|
-
|
|
14
|
-
namespace signalsmith {
|
|
15
|
-
namespace spectral {
|
|
16
|
-
/** @defgroup Spectral Spectral Processing
|
|
17
|
-
@brief Tools for frequency-domain manipulation of audio signals
|
|
18
|
-
|
|
19
|
-
@{
|
|
20
|
-
@file
|
|
21
|
-
*/
|
|
22
|
-
|
|
23
|
-
/** @brief An FFT with built-in windowing and round-trip scaling
|
|
24
|
-
|
|
25
|
-
This uses a Modified Real FFT, which applies half-bin shift before the transform. The result therefore has `N/2` bins, centred at the frequencies: `(i + 0.5)/N`.
|
|
26
|
-
|
|
27
|
-
This avoids the awkward (real-valued) bands for DC-offset and Nyquist.
|
|
28
|
-
*/
|
|
29
|
-
template<typename Sample>
|
|
30
|
-
class WindowedFFT {
|
|
31
|
-
using MRFFT = signalsmith::fft::ModifiedRealFFT<Sample>;
|
|
32
|
-
using Complex = std::complex<Sample>;
|
|
33
|
-
MRFFT mrfft{2};
|
|
34
|
-
|
|
35
|
-
std::vector<Sample> fftWindow;
|
|
36
|
-
std::vector<Sample> timeBuffer;
|
|
37
|
-
int offsetSamples = 0;
|
|
38
|
-
public:
|
|
39
|
-
/// Returns a fast FFT size <= `size`
|
|
40
|
-
static int fastSizeAbove(int size, int divisor=1) {
|
|
41
|
-
return MRFFT::fastSizeAbove(size/divisor)*divisor;
|
|
42
|
-
}
|
|
43
|
-
/// Returns a fast FFT size >= `size`
|
|
44
|
-
static int fastSizeBelow(int size, int divisor=1) {
|
|
45
|
-
return MRFFT::fastSizeBelow(1 + (size - 1)/divisor)*divisor;
|
|
46
|
-
}
|
|
47
|
-
|
|
48
|
-
WindowedFFT() {}
|
|
49
|
-
WindowedFFT(int size, int rotateSamples=0) {
|
|
50
|
-
setSize(size, rotateSamples);
|
|
51
|
-
}
|
|
52
|
-
template<class WindowFn>
|
|
53
|
-
WindowedFFT(int size, WindowFn fn, Sample windowOffset=0.5, int rotateSamples=0) {
|
|
54
|
-
setSize(size, fn, windowOffset, rotateSamples);
|
|
55
|
-
}
|
|
56
|
-
|
|
57
|
-
/// Sets the size, returning the window for modification (initially all 1s)
|
|
58
|
-
std::vector<Sample> & setSizeWindow(int size, int rotateSamples=0) {
|
|
59
|
-
mrfft.setSize(size);
|
|
60
|
-
fftWindow.assign(size, 1);
|
|
61
|
-
timeBuffer.resize(size);
|
|
62
|
-
offsetSamples = rotateSamples;
|
|
63
|
-
if (offsetSamples < 0) offsetSamples += size; // TODO: for a negative rotation, the other half of the result is inverted
|
|
64
|
-
return fftWindow;
|
|
65
|
-
}
|
|
66
|
-
/// Sets the FFT size, with a user-defined functor for the window
|
|
67
|
-
template<class WindowFn>
|
|
68
|
-
void setSize(int size, WindowFn fn, Sample windowOffset=0.5, int rotateSamples=0) {
|
|
69
|
-
setSizeWindow(size, rotateSamples);
|
|
70
|
-
|
|
71
|
-
Sample invSize = 1/(Sample)size;
|
|
72
|
-
for (int i = 0; i < size; ++i) {
|
|
73
|
-
Sample r = (i + windowOffset)*invSize;
|
|
74
|
-
fftWindow[i] = fn(r);
|
|
75
|
-
}
|
|
76
|
-
}
|
|
77
|
-
/// Sets the size (using the default Blackman-Harris window)
|
|
78
|
-
void setSize(int size, int rotateSamples=0) {
|
|
79
|
-
setSize(size, [](double x) {
|
|
80
|
-
double phase = 2 * audioapi::PI * x;
|
|
81
|
-
// Blackman-Harris
|
|
82
|
-
return 0.35875 - 0.48829*std::cos(phase) + 0.14128*std::cos(phase*2) - 0.01168*std::cos(phase*3);
|
|
83
|
-
}, Sample(0.5), rotateSamples);
|
|
84
|
-
}
|
|
85
|
-
|
|
86
|
-
const std::vector<Sample> & window() const {
|
|
87
|
-
return this->fftWindow;
|
|
88
|
-
}
|
|
89
|
-
int size() const {
|
|
90
|
-
return mrfft.size();
|
|
91
|
-
}
|
|
92
|
-
|
|
93
|
-
/// Performs an FFT, with windowing and rotation (if enabled)
|
|
94
|
-
template<bool withWindow=true, bool withScaling=false, class Input, class Output>
|
|
95
|
-
void fft(Input &&input, Output &&output) {
|
|
96
|
-
int fftSize = size();
|
|
97
|
-
const Sample norm = (withScaling ? 1/(Sample)fftSize : 1);
|
|
98
|
-
for (int i = 0; i < offsetSamples; ++i) {
|
|
99
|
-
// Inverted polarity since we're using the MRFFT
|
|
100
|
-
timeBuffer[i + fftSize - offsetSamples] = -input[i]*norm*(withWindow ? fftWindow[i] : Sample(1));
|
|
101
|
-
}
|
|
102
|
-
for (int i = offsetSamples; i < fftSize; ++i) {
|
|
103
|
-
timeBuffer[i - offsetSamples] = input[i]*norm*(withWindow ? fftWindow[i] : Sample(1));
|
|
104
|
-
}
|
|
105
|
-
mrfft.fft(timeBuffer, output);
|
|
106
|
-
}
|
|
107
|
-
/// Performs an FFT (no windowing or rotation)
|
|
108
|
-
template<class Input, class Output>
|
|
109
|
-
void fftRaw(Input &&input, Output &&output) {
|
|
110
|
-
mrfft.fft(input, output);
|
|
111
|
-
}
|
|
112
|
-
|
|
113
|
-
/// Inverse FFT, with windowing, 1/N scaling and rotation (if enabled)
|
|
114
|
-
template<bool withWindow=true, bool withScaling=true, class Input, class Output>
|
|
115
|
-
void ifft(Input &&input, Output &&output) {
|
|
116
|
-
mrfft.ifft(input, timeBuffer);
|
|
117
|
-
int fftSize = mrfft.size();
|
|
118
|
-
const Sample norm = (withScaling ? 1/(Sample)fftSize : 1);
|
|
119
|
-
|
|
120
|
-
for (int i = 0; i < offsetSamples; ++i) {
|
|
121
|
-
// Inverted polarity since we're using the MRFFT
|
|
122
|
-
output[i] = -timeBuffer[i + fftSize - offsetSamples]*norm*(withWindow ? fftWindow[i] : Sample(1));
|
|
123
|
-
}
|
|
124
|
-
for (int i = offsetSamples; i < fftSize; ++i) {
|
|
125
|
-
output[i] = timeBuffer[i - offsetSamples]*norm*(withWindow ? fftWindow[i] : Sample(1));
|
|
126
|
-
}
|
|
127
|
-
}
|
|
128
|
-
/// Performs an IFFT (no windowing, scaling or rotation)
|
|
129
|
-
template<class Input, class Output>
|
|
130
|
-
void ifftRaw(Input &&input, Output &&output) {
|
|
131
|
-
mrfft.ifft(input, output);
|
|
132
|
-
}
|
|
133
|
-
};
|
|
134
|
-
|
|
135
|
-
/** STFT synthesis, built on a `MultiBuffer`.
|
|
136
|
-
|
|
137
|
-
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.
|
|
138
|
-
|
|
139
|
-
\diagram{stft-aliasing-simulated.svg,Simulated bad-case aliasing (random phase-shift for each band) for overlapping ratios}
|
|
140
|
-
|
|
141
|
-
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.
|
|
142
|
-
|
|
143
|
-
\diagram{stft-buffer-validity.svg}
|
|
144
|
-
|
|
145
|
-
You move the valid index along using `.ensureValid()`, passing in a functor which provides spectra (using `.analyse()` and/or direct modification through `.spectrum[c]`):
|
|
146
|
-
|
|
147
|
-
\code
|
|
148
|
-
void processSample(...) {
|
|
149
|
-
stft.ensureValid([&](int) {
|
|
150
|
-
// Here, we introduce (1 - windowSize) of latency
|
|
151
|
-
stft.analyse(inputBuffer.view(1 - windowSize))
|
|
152
|
-
});
|
|
153
|
-
// read as a MultiBuffer
|
|
154
|
-
auto result = stft.at(0);
|
|
155
|
-
++stft; // also moves the latest valid index
|
|
156
|
-
}
|
|
157
|
-
|
|
158
|
-
void processBlock(...) {
|
|
159
|
-
// assuming `historyLength` == blockSize
|
|
160
|
-
stft.ensureValid(blockSize, [&](int blockStartIndex) {
|
|
161
|
-
int inputStart = blockStartIndex + (1 - windowSize);
|
|
162
|
-
stft.analyse(inputBuffer.view(inputStart));
|
|
163
|
-
});
|
|
164
|
-
auto earliestValid = stft.at(0);
|
|
165
|
-
auto latestValid = stft.at(blockSize);
|
|
166
|
-
stft += blockSize;
|
|
167
|
-
}
|
|
168
|
-
\endcode
|
|
169
|
-
|
|
170
|
-
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`.
|
|
171
|
-
*/
|
|
172
|
-
template<typename Sample>
|
|
173
|
-
class STFT : public signalsmith::delay::MultiBuffer<Sample> {
|
|
174
|
-
|
|
175
|
-
using Super = signalsmith::delay::MultiBuffer<Sample>;
|
|
176
|
-
using Complex = std::complex<Sample>;
|
|
177
|
-
|
|
178
|
-
int channels = 0, _windowSize = 0, _fftSize = 0, _interval = 1;
|
|
179
|
-
int validUntilIndex = 0;
|
|
180
|
-
|
|
181
|
-
class MultiSpectrum {
|
|
182
|
-
int channels, stride;
|
|
183
|
-
std::vector<Complex> buffer;
|
|
184
|
-
public:
|
|
185
|
-
MultiSpectrum() : MultiSpectrum(0, 0) {}
|
|
186
|
-
MultiSpectrum(int channels, int bands) : channels(channels), stride(bands), buffer(channels*bands, 0) {}
|
|
187
|
-
|
|
188
|
-
void resize(int nChannels, int nBands) {
|
|
189
|
-
channels = nChannels;
|
|
190
|
-
stride = nBands;
|
|
191
|
-
buffer.assign(channels*stride, 0);
|
|
192
|
-
}
|
|
193
|
-
|
|
194
|
-
void reset() {
|
|
195
|
-
buffer.assign(buffer.size(), 0);
|
|
196
|
-
}
|
|
197
|
-
|
|
198
|
-
void swap(MultiSpectrum &other) {
|
|
199
|
-
using std::swap;
|
|
200
|
-
swap(buffer, other.buffer);
|
|
201
|
-
}
|
|
202
|
-
|
|
203
|
-
Complex * operator [](int channel) {
|
|
204
|
-
return buffer.data() + channel*stride;
|
|
205
|
-
}
|
|
206
|
-
const Complex * operator [](int channel) const {
|
|
207
|
-
return buffer.data() + channel*stride;
|
|
208
|
-
}
|
|
209
|
-
};
|
|
210
|
-
std::vector<Sample> timeBuffer;
|
|
211
|
-
|
|
212
|
-
bool rotate = false;
|
|
213
|
-
void resizeInternal(int newChannels, int windowSize, int newInterval, int historyLength, int zeroPadding) {
|
|
214
|
-
Super::resize(newChannels,
|
|
215
|
-
windowSize /* for output summing */
|
|
216
|
-
+ newInterval /* so we can read `windowSize` ahead (we'll be at most `interval-1` from the most recent block */
|
|
217
|
-
+ historyLength);
|
|
218
|
-
|
|
219
|
-
int fftSize = fft.fastSizeAbove(windowSize + zeroPadding);
|
|
220
|
-
|
|
221
|
-
this->channels = newChannels;
|
|
222
|
-
_windowSize = windowSize;
|
|
223
|
-
this->_fftSize = fftSize;
|
|
224
|
-
this->_interval = newInterval;
|
|
225
|
-
validUntilIndex = -1;
|
|
226
|
-
|
|
227
|
-
setWindow(windowShape, rotate);
|
|
228
|
-
|
|
229
|
-
spectrum.resize(channels, fftSize/2);
|
|
230
|
-
timeBuffer.resize(fftSize);
|
|
231
|
-
}
|
|
232
|
-
public:
|
|
233
|
-
enum class Window {kaiser, acg};
|
|
234
|
-
/// \deprecated use `.setWindow()` which actually updates the window when you change it
|
|
235
|
-
Window windowShape = Window::kaiser;
|
|
236
|
-
// for convenience
|
|
237
|
-
static constexpr Window kaiser = Window::kaiser;
|
|
238
|
-
static constexpr Window acg = Window::acg;
|
|
239
|
-
|
|
240
|
-
/** Swaps between the default (Kaiser) shape and Approximate Confined Gaussian (ACG).
|
|
241
|
-
\diagram{stft-windows.svg,Default (Kaiser) windows and partial cumulative sum}
|
|
242
|
-
The ACG has better rolloff since its edges go to 0:
|
|
243
|
-
\diagram{stft-windows-acg.svg,ACG windows and partial cumulative sum}
|
|
244
|
-
However, it generally has worse performance in terms of total sidelobe energy, affecting worst-case aliasing levels for (most) higher overlap ratios:
|
|
245
|
-
\diagram{stft-aliasing-simulated-acg.svg,Simulated bad-case aliasing for ACG windows - compare with above}*/
|
|
246
|
-
// TODO: these should both be set before resize()
|
|
247
|
-
void setWindow(Window shape, bool rotateToZero=false) {
|
|
248
|
-
windowShape = shape;
|
|
249
|
-
rotate = rotateToZero;
|
|
250
|
-
|
|
251
|
-
auto &window = fft.setSizeWindow(_fftSize, rotateToZero ? _windowSize/2 : 0);
|
|
252
|
-
if (windowShape == Window::kaiser) {
|
|
253
|
-
/// Roughly optimal Kaiser for STFT analysis (forced to perfect reconstruction)
|
|
254
|
-
audioapi::dsp::Kaiser::withBandwidth(_windowSize/_interval, true).apply(window.data(), _windowSize);
|
|
255
|
-
} else {
|
|
256
|
-
audioapi::dsp::ApproximateConfinedGaussian::withBandwidth(_windowSize/_interval).apply(window.data(), _windowSize);
|
|
257
|
-
}
|
|
258
|
-
audioapi::dsp::WindowFunction::forcePerfectReconstruction(window.data(), _windowSize, _interval);
|
|
259
|
-
|
|
260
|
-
// TODO: fill extra bits of an input buffer with NaN/Infinity, to break this, and then fix by adding zero-padding to WindowedFFT (as opposed to zero-valued window sections)
|
|
261
|
-
for (int i = _windowSize; i < _fftSize; ++i) {
|
|
262
|
-
window[i] = 0;
|
|
263
|
-
}
|
|
264
|
-
}
|
|
265
|
-
|
|
266
|
-
using Spectrum = MultiSpectrum;
|
|
267
|
-
Spectrum spectrum;
|
|
268
|
-
WindowedFFT<Sample> fft;
|
|
269
|
-
|
|
270
|
-
STFT() {}
|
|
271
|
-
/// Parameters passed straight to `.resize()`
|
|
272
|
-
STFT(int channels, int windowSize, int interval, int historyLength=0, int zeroPadding=0) {
|
|
273
|
-
resize(channels, windowSize, interval, historyLength, zeroPadding);
|
|
274
|
-
}
|
|
275
|
-
|
|
276
|
-
/// Sets the channel-count, FFT size and interval.
|
|
277
|
-
void resize(int nChannels, int windowSize, int interval, int historyLength=0, int zeroPadding=0) {
|
|
278
|
-
resizeInternal(nChannels, windowSize, interval, historyLength, zeroPadding);
|
|
279
|
-
}
|
|
280
|
-
|
|
281
|
-
int windowSize() const {
|
|
282
|
-
return _windowSize;
|
|
283
|
-
}
|
|
284
|
-
int fftSize() const {
|
|
285
|
-
return _fftSize;
|
|
286
|
-
}
|
|
287
|
-
int interval() const {
|
|
288
|
-
return _interval;
|
|
289
|
-
}
|
|
290
|
-
/// Returns the (analysis and synthesis) window
|
|
291
|
-
decltype(fft.window()) window() const {
|
|
292
|
-
return fft.window();
|
|
293
|
-
}
|
|
294
|
-
/// Calculates the effective window for the partially-summed future output (relative to the most recent block)
|
|
295
|
-
std::vector<Sample> partialSumWindow(bool includeLatestBlock=true) const {
|
|
296
|
-
const auto &w = window();
|
|
297
|
-
std::vector<Sample> result(_windowSize, 0);
|
|
298
|
-
int firstOffset = (includeLatestBlock ? 0 : _interval);
|
|
299
|
-
for (int offset = firstOffset; offset < _windowSize; offset += _interval) {
|
|
300
|
-
for (int i = 0; i < _windowSize - offset; ++i) {
|
|
301
|
-
Sample value = w[i + offset];
|
|
302
|
-
result[i] += value*value;
|
|
303
|
-
}
|
|
304
|
-
}
|
|
305
|
-
return result;
|
|
306
|
-
}
|
|
307
|
-
|
|
308
|
-
/// Resets everything - since we clear the output sum, it will take `windowSize` samples to get proper output.
|
|
309
|
-
void reset() {
|
|
310
|
-
Super::reset();
|
|
311
|
-
spectrum.reset();
|
|
312
|
-
validUntilIndex = -1;
|
|
313
|
-
}
|
|
314
|
-
|
|
315
|
-
/** Generates valid output up to the specified index (or 0), using the callback as many times as needed.
|
|
316
|
-
|
|
317
|
-
The callback should be a functor accepting a single integer argument, which is the index for which a spectrum is required.
|
|
318
|
-
|
|
319
|
-
The block created from these spectra will start at this index in the output, plus `.latency()`.
|
|
320
|
-
*/
|
|
321
|
-
template<class AnalysisFn>
|
|
322
|
-
void ensureValid(int i, AnalysisFn fn) {
|
|
323
|
-
while (validUntilIndex < i) {
|
|
324
|
-
int blockIndex = validUntilIndex + 1;
|
|
325
|
-
fn(blockIndex);
|
|
326
|
-
|
|
327
|
-
auto output = this->view(blockIndex);
|
|
328
|
-
for (int c = 0; c < channels; ++c) {
|
|
329
|
-
auto channel = output[c];
|
|
330
|
-
|
|
331
|
-
// Clear out the future sum, a window-length and an interval ahead
|
|
332
|
-
for (int wi = _windowSize; wi < _windowSize + _interval; ++wi) {
|
|
333
|
-
channel[wi] = 0;
|
|
334
|
-
}
|
|
335
|
-
|
|
336
|
-
// Add in the IFFT'd result
|
|
337
|
-
fft.ifft(spectrum[c], timeBuffer);
|
|
338
|
-
for (int wi = 0; wi < _windowSize; ++wi) {
|
|
339
|
-
channel[wi] += timeBuffer[wi];
|
|
340
|
-
}
|
|
341
|
-
}
|
|
342
|
-
validUntilIndex += _interval;
|
|
343
|
-
}
|
|
344
|
-
}
|
|
345
|
-
/// The same as above, assuming index 0
|
|
346
|
-
template<class AnalysisFn>
|
|
347
|
-
void ensureValid(AnalysisFn fn) {
|
|
348
|
-
return ensureValid(0, fn);
|
|
349
|
-
}
|
|
350
|
-
/// Returns the next invalid index (a.k.a. the index of the next block)
|
|
351
|
-
int nextInvalid() const {
|
|
352
|
-
return validUntilIndex + 1;
|
|
353
|
-
}
|
|
354
|
-
|
|
355
|
-
/** Analyse a multi-channel input, for any type where `data[channel][index]` returns samples
|
|
356
|
-
|
|
357
|
-
Results can be read/edited using `.spectrum`. */
|
|
358
|
-
template<class Data>
|
|
359
|
-
void analyse(Data &&data) {
|
|
360
|
-
for (int c = 0; c < channels; ++c) {
|
|
361
|
-
fft.fft(data[c], spectrum[c]);
|
|
362
|
-
}
|
|
363
|
-
}
|
|
364
|
-
template<class Data>
|
|
365
|
-
void analyse(int c, Data &&data) {
|
|
366
|
-
fft.fft(data, spectrum[c]);
|
|
367
|
-
}
|
|
368
|
-
/// Analyse without windowing or zero-rotation
|
|
369
|
-
template<class Data>
|
|
370
|
-
void analyseRaw(Data &&data) {
|
|
371
|
-
for (int c = 0; c < channels; ++c) {
|
|
372
|
-
fft.fftRaw(data[c], spectrum[c]);
|
|
373
|
-
}
|
|
374
|
-
}
|
|
375
|
-
template<class Data>
|
|
376
|
-
void analyseRaw(int c, Data &&data) {
|
|
377
|
-
fft.fftRaw(data, spectrum[c]);
|
|
378
|
-
}
|
|
379
|
-
|
|
380
|
-
int bands() const {
|
|
381
|
-
return _fftSize/2;
|
|
382
|
-
}
|
|
383
|
-
|
|
384
|
-
/** Internal latency (between the block-index requested in `.ensureValid()` and its position in the output)
|
|
385
|
-
|
|
386
|
-
Currently unused, but it's in here to allow for a future implementation which spreads the FFT calculations out across each interval.*/
|
|
387
|
-
int latency() {
|
|
388
|
-
return 0;
|
|
389
|
-
}
|
|
390
|
-
|
|
391
|
-
// @name Shift the underlying buffer (moving the "valid" index accordingly)
|
|
392
|
-
// @{
|
|
393
|
-
STFT & operator ++() {
|
|
394
|
-
Super::operator ++();
|
|
395
|
-
validUntilIndex--;
|
|
396
|
-
return *this;
|
|
397
|
-
}
|
|
398
|
-
STFT & operator +=(int i) {
|
|
399
|
-
Super::operator +=(i);
|
|
400
|
-
validUntilIndex -= i;
|
|
401
|
-
return *this;
|
|
402
|
-
}
|
|
403
|
-
STFT & operator --() {
|
|
404
|
-
Super::operator --();
|
|
405
|
-
validUntilIndex++;
|
|
406
|
-
return *this;
|
|
407
|
-
}
|
|
408
|
-
STFT & operator -=(int i) {
|
|
409
|
-
Super::operator -=(i);
|
|
410
|
-
validUntilIndex += i;
|
|
411
|
-
return *this;
|
|
412
|
-
}
|
|
413
|
-
// @}
|
|
414
|
-
|
|
415
|
-
typename Super::MutableView operator ++(int postIncrement) {
|
|
416
|
-
auto result = Super::operator ++(postIncrement);
|
|
417
|
-
validUntilIndex--;
|
|
418
|
-
return result;
|
|
419
|
-
}
|
|
420
|
-
typename Super::MutableView operator --(int postIncrement) {
|
|
421
|
-
auto result = Super::operator --(postIncrement);
|
|
422
|
-
validUntilIndex++;
|
|
423
|
-
return result;
|
|
424
|
-
}
|
|
425
|
-
};
|
|
426
|
-
|
|
427
|
-
/** STFT processing, with input/output.
|
|
428
|
-
Before calling `.ensureValid(index)`, you should make sure the input is filled up to `index`.
|
|
429
|
-
*/
|
|
430
|
-
template<typename Sample>
|
|
431
|
-
class ProcessSTFT : public STFT<Sample> {
|
|
432
|
-
using Super = STFT<Sample>;
|
|
433
|
-
public:
|
|
434
|
-
signalsmith::delay::MultiBuffer<Sample> input;
|
|
435
|
-
|
|
436
|
-
ProcessSTFT(int inChannels, int outChannels, int windowSize, int interval, int historyLength=0) {
|
|
437
|
-
resize(inChannels, outChannels, windowSize, interval, historyLength);
|
|
438
|
-
}
|
|
439
|
-
|
|
440
|
-
/** Alter the spectrum, using input up to this point, for the output block starting from this point.
|
|
441
|
-
Sub-classes should replace this with whatever processing is desired. */
|
|
442
|
-
virtual void processSpectrum(int /*blockIndex*/) {}
|
|
443
|
-
|
|
444
|
-
/// Sets the input/output channels, FFT size and interval.
|
|
445
|
-
void resize(int inChannels, int outChannels, int windowSize, int interval, int historyLength=0) {
|
|
446
|
-
Super::resize(outChannels, windowSize, interval, historyLength);
|
|
447
|
-
input.resize(inChannels, windowSize + interval + historyLength);
|
|
448
|
-
}
|
|
449
|
-
void reset(Sample value=Sample()) {
|
|
450
|
-
Super::reset(value);
|
|
451
|
-
input.reset(value);
|
|
452
|
-
}
|
|
453
|
-
|
|
454
|
-
/// Internal latency, including buffering samples for analysis.
|
|
455
|
-
int latency() {
|
|
456
|
-
return Super::latency() + (this->windowSize() - 1);
|
|
457
|
-
}
|
|
458
|
-
|
|
459
|
-
void ensureValid(int i=0) {
|
|
460
|
-
Super::ensureValid(i, [&](int blockIndex) {
|
|
461
|
-
this->analyse(input.view(blockIndex - this->windowSize() + 1));
|
|
462
|
-
this->processSpectrum(blockIndex);
|
|
463
|
-
});
|
|
464
|
-
}
|
|
465
|
-
|
|
466
|
-
// @name Shift the output, input, and valid index.
|
|
467
|
-
// @{
|
|
468
|
-
ProcessSTFT & operator ++() {
|
|
469
|
-
Super::operator ++();
|
|
470
|
-
++input;
|
|
471
|
-
return *this;
|
|
472
|
-
}
|
|
473
|
-
ProcessSTFT & operator +=(int i) {
|
|
474
|
-
Super::operator +=(i);
|
|
475
|
-
input += i;
|
|
476
|
-
return *this;
|
|
477
|
-
}
|
|
478
|
-
ProcessSTFT & operator --() {
|
|
479
|
-
Super::operator --();
|
|
480
|
-
--input;
|
|
481
|
-
return *this;
|
|
482
|
-
}
|
|
483
|
-
ProcessSTFT & operator -=(int i) {
|
|
484
|
-
Super::operator -=(i);
|
|
485
|
-
input -= i;
|
|
486
|
-
return *this;
|
|
487
|
-
}
|
|
488
|
-
// @}
|
|
489
|
-
};
|
|
490
|
-
|
|
491
|
-
/** @} */
|
|
492
|
-
}} // signalsmith::spectral::
|
|
493
|
-
#endif // include guard
|