1054 lines
40 KiB
C++
1054 lines
40 KiB
C++
/**********************************************************************
|
|
|
|
Audacity: A Digital Audio Editor
|
|
|
|
VoiceKey.cpp
|
|
|
|
?? Dominic Mazzoni
|
|
?? Shane Muller
|
|
|
|
*******************************************************************//*!
|
|
|
|
\class VoiceKey
|
|
|
|
\brief
|
|
This implements a voice key, detecting either the next "ON"
|
|
or "OFF" point
|
|
|
|
*//*******************************************************************/
|
|
|
|
|
|
#include "Audacity.h"
|
|
|
|
#include "VoiceKey.h"
|
|
#include <wx/string.h>
|
|
#include <wx/msgdlg.h>
|
|
#include <math.h>
|
|
#include <stdio.h>
|
|
|
|
#include <wx/filedlg.h>
|
|
#include <wx/msgdlg.h>
|
|
#include <wx/textfile.h>
|
|
#include <wx/intl.h>
|
|
#include <iostream>
|
|
|
|
using std::cout;
|
|
using std::endl;
|
|
|
|
|
|
|
|
VoiceKey::VoiceKey(){
|
|
|
|
mWindowSize = 0.01; //size of analysis window in seconds
|
|
|
|
mEnergyMean = .0006; // reasonable initial levels assuming sampling rate of
|
|
mEnergySD = .0002; // 44100 hertz
|
|
mSignChangesMean = .08;
|
|
mSignChangesSD= .02;
|
|
mDirectionChangesMean = .25;
|
|
mDirectionChangesSD = .2;
|
|
|
|
AdjustThreshold(2);
|
|
|
|
mSilentWindowSize = .05; //Amount of time (in seconds) below threshold to call it silence
|
|
mSignalWindowSize = .05; //Amount of time (in seconds) above threshold to call it signal
|
|
|
|
|
|
mUseEnergy = true;
|
|
mUseSignChangesLow = false;
|
|
mUseSignChangesHigh = false;
|
|
mUseDirectionChangesLow = false;
|
|
mUseDirectionChangesHigh = false;
|
|
|
|
|
|
};
|
|
|
|
|
|
VoiceKey::~VoiceKey(){
|
|
};
|
|
|
|
|
|
|
|
//---------------------------------------------------------------------------
|
|
// VoiceKey::On/Off Forward/Backward
|
|
// This operates in two phases:
|
|
// First, you take chunks of samples that are WindowSize big.
|
|
// If you have a run of them where something passes the threshold for SignalWindowSize seconds,
|
|
// you return to the last empty block and scan forward one sample at a time until you find the
|
|
// starting point of the speech.
|
|
|
|
|
|
|
|
|
|
//Move forward to find an ON region.
|
|
sampleCount VoiceKey::OnForward (WaveTrack & t, sampleCount start, sampleCount len) {
|
|
|
|
if((mWindowSize) >= len+10){
|
|
|
|
wxMessageBox(_("Selection is too small to use voice key."));
|
|
return start;
|
|
}
|
|
else{
|
|
|
|
sampleCount lastsubthresholdsample; // keeps track of the sample number of the last sample to not exceed the threshold
|
|
|
|
//Change the millisecond-based parameters into sample-based parameters
|
|
double rate = t.GetRate(); //Translates seconds to samples
|
|
unsigned int WindowSizeInt = (unsigned int)(rate * mWindowSize); //Size of window to examine
|
|
unsigned int SignalWindowSizeInt = (unsigned int)(rate * mSignalWindowSize); //This much signal is necessary to trip key
|
|
|
|
int samplesleft = len - WindowSizeInt; //Indexes the number of samples remaining in the selection
|
|
lastsubthresholdsample = start; //start this off at the selection start
|
|
unsigned int i; //iterates through waveblock
|
|
int blockruns=0; //keeps track of the number of consecutive above-threshold blocks
|
|
int blocksize; //The final block may be smaller than WindowSizeInt, so use this
|
|
|
|
|
|
//This loop goes through the selection a block at a time. If a long enough run
|
|
//of above-threshold blocks occur, we return to the last sub-threshold block and
|
|
//go through one sample at a time.
|
|
//If there are fewer than 10 samples leftover, don't bother.
|
|
|
|
for(i = start; samplesleft >=10; i+=(WindowSizeInt-1) , samplesleft -= (WindowSizeInt -1)){
|
|
|
|
//Set blocksize so that it is the right size
|
|
if((unsigned int)samplesleft < WindowSizeInt){
|
|
blocksize = samplesleft;
|
|
}
|
|
else{
|
|
blocksize = WindowSizeInt;
|
|
}
|
|
|
|
//Test whether we are above threshold (the number of stats)
|
|
if(AboveThreshold(t,i,blocksize))
|
|
{
|
|
blockruns++; //Hit
|
|
} else{
|
|
blockruns=0; //Miss--start over
|
|
lastsubthresholdsample = i;
|
|
}
|
|
|
|
//If the blockrun is long enough, break out of the loop early:
|
|
if(blockruns > mSignalWindowSize/mWindowSize)
|
|
break;
|
|
|
|
}
|
|
|
|
//Now, if we broke out early (samplesleft > 10), go back to the lastsubthresholdsample and look more carefully
|
|
if(samplesleft > 10){
|
|
|
|
|
|
//Calculate how many to scan through--we only have to go through (at most)
|
|
//the first window + 1 samples--but we need another window samples to draw from.
|
|
samplesleft = 2*WindowSizeInt+1;
|
|
|
|
//To speed things up, create a local buffer to store things in, to avoid the costly t.Get();
|
|
//Only go through the first SignalWindowSizeInt samples, and choose the first that trips the key.
|
|
sampleFormat *buffer = new sampleFormat[samplesleft];
|
|
t.Get((samplePtr)buffer, floatSample,lastsubthresholdsample,samplesleft);
|
|
|
|
|
|
|
|
//Initialize these trend markers atrend and ztrend. They keep track of the
|
|
//up/down trends at the start and end of the evaluation window.
|
|
int atrend = sgn(buffer[1]-buffer[0]);
|
|
int ztrend = sgn(buffer[WindowSizeInt+1]-buffer[WindowSizeInt]);
|
|
|
|
|
|
double erg=0;
|
|
double sc=0;
|
|
double dc=0;
|
|
|
|
//Get initial test statistic values.
|
|
if(mUseEnergy)
|
|
erg = TestEnergy(t, lastsubthresholdsample, WindowSizeInt);
|
|
|
|
if(mUseSignChangesLow || mUseSignChangesHigh)
|
|
sc = TestSignChanges(t,lastsubthresholdsample, WindowSizeInt);
|
|
|
|
if(mUseDirectionChangesLow || mUseDirectionChangesHigh)
|
|
dc = TestDirectionChanges(t,lastsubthresholdsample,WindowSizeInt);
|
|
|
|
|
|
//Now, go through the sound again, sample by sample.
|
|
for(i=0; i<SignalWindowSizeInt-WindowSizeInt;i++){
|
|
|
|
int tests = 0;
|
|
int testThreshold = 0;
|
|
//Update the test statistics
|
|
if(mUseEnergy)
|
|
{
|
|
TestEnergyUpdate(erg, WindowSizeInt,buffer[i],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(erg>mThresholdEnergy);
|
|
testThreshold++;
|
|
}
|
|
if(mUseSignChangesLow)
|
|
{
|
|
TestSignChangesUpdate(sc,WindowSizeInt,buffer[i],buffer[i+1],buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(sc < mThresholdSignChangesLower);
|
|
testThreshold++;
|
|
}
|
|
|
|
if(mUseSignChangesHigh)
|
|
{
|
|
TestSignChangesUpdate(sc,WindowSizeInt,buffer[i],buffer[i+1],buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(sc > mThresholdSignChangesUpper);
|
|
testThreshold++;
|
|
}
|
|
|
|
if(mUseDirectionChangesLow)
|
|
{
|
|
TestDirectionChangesUpdate(dc,WindowSizeInt,atrend,buffer[i],buffer[i+1],ztrend,buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(dc < mThresholdDirectionChangesLower);
|
|
testThreshold++;
|
|
}
|
|
|
|
if(mUseDirectionChangesHigh)
|
|
{
|
|
TestDirectionChangesUpdate(dc,WindowSizeInt,atrend,buffer[i],buffer[i+1],ztrend,buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(dc > mThresholdDirectionChangesUpper);
|
|
testThreshold++;
|
|
}
|
|
|
|
|
|
|
|
if(tests >= testThreshold)
|
|
{ //Finish off on the first hit
|
|
break;
|
|
}
|
|
}
|
|
|
|
//When we get here, i+lastsubthresholdsample is the best guess for where the word starts
|
|
delete [] buffer;
|
|
return i+lastsubthresholdsample;
|
|
}
|
|
else{
|
|
//If we failed to find anything, return the start position
|
|
return start ;
|
|
}
|
|
}
|
|
}
|
|
|
|
//Move backward from end to find an ON region.
|
|
sampleCount VoiceKey::OnBackward (WaveTrack & t, sampleCount end, sampleCount len) {
|
|
|
|
|
|
if((mWindowSize) >= len+10){
|
|
|
|
wxMessageBox(_("Selection is too small to use voice key."));
|
|
return end;
|
|
}
|
|
else{
|
|
|
|
sampleCount lastsubthresholdsample; // keeps track of the sample number of the last sample to not exceed the threshold
|
|
|
|
//Change the millisecond-based parameters into sample-based parameters
|
|
double rate = t.GetRate(); //Translates seconds to samples
|
|
unsigned int WindowSizeInt = (unsigned int)(rate * mWindowSize); //Size of window to examine
|
|
//unsigned int SilentWindowSizeInt = (unsigned int)(rate * mSilentWindowSize); //This much signal is necessary to trip key
|
|
|
|
int samplesleft = len - WindowSizeInt; //Indexes the number of samples remaining in the selection
|
|
lastsubthresholdsample = end; //start this off at the end
|
|
unsigned int i; //iterates through waveblock
|
|
int blockruns=0; //keeps track of the number of consecutive above-threshold blocks
|
|
int blocksize; //The final block may be smaller than WindowSizeInt, so use this
|
|
|
|
|
|
//This loop goes through the selection a block at a time in reverse order. If a long enough run
|
|
//of above-threshold blocks occur, we return to the last sub-threshold block and
|
|
//go through one sample at a time.
|
|
//If there are fewer than 10 samples leftover, don't bother.
|
|
for(i = end - WindowSizeInt; samplesleft >=10; i-=(WindowSizeInt-1) , samplesleft -= (WindowSizeInt -1)){
|
|
|
|
//Set blocksize so that it is the right size
|
|
if(samplesleft < (int)WindowSizeInt){
|
|
blocksize = samplesleft;
|
|
}
|
|
else{
|
|
blocksize = WindowSizeInt;
|
|
}
|
|
|
|
|
|
//Test whether we are above threshold
|
|
if(AboveThreshold(t,i,blocksize))
|
|
{
|
|
blockruns++; //Hit
|
|
}
|
|
else
|
|
{
|
|
blockruns=0; //Miss--start over
|
|
lastsubthresholdsample = i+WindowSizeInt;
|
|
}
|
|
|
|
//If the blockrun is long enough, break out of the loop early:
|
|
if(blockruns > mSilentWindowSize/mWindowSize)
|
|
break;
|
|
|
|
}
|
|
|
|
//Now, if we broke out early (samplesleft > 10), go back to the lastsubthresholdsample and look more carefully
|
|
if(samplesleft > 10){
|
|
|
|
//Calculate how many to scan through--we only have to go through (at most)
|
|
//the first window + 1 samples--but we need another window samples to draw from.
|
|
samplesleft = 2*WindowSizeInt+1;
|
|
|
|
//To speed things up, create a local buffer to store things in, to avoid the costly t.Get();
|
|
//Only go through the first mSilentWindowSizeInt samples, and choose the first that trips the key.
|
|
sampleFormat *buffer = new sampleFormat[samplesleft];
|
|
t.Get((samplePtr)buffer, floatSample, lastsubthresholdsample-samplesleft,samplesleft);
|
|
|
|
//Initialize these trend markers atrend and ztrend. They keep track of the
|
|
//up/down trends at the start and end of the evaluation window.
|
|
int atrend = sgn(buffer[samplesleft - 2]-buffer[samplesleft - 1]);
|
|
int ztrend = sgn(buffer[samplesleft - WindowSizeInt-2]-buffer[samplesleft - WindowSizeInt-2]);
|
|
|
|
double erg=0;
|
|
double sc = 0;
|
|
double dc = 0;
|
|
|
|
//Get initial test statistic values.
|
|
if(mUseEnergy)
|
|
erg = TestEnergy(t, lastsubthresholdsample, WindowSizeInt);
|
|
if(mUseSignChangesLow || mUseSignChangesHigh)
|
|
sc = TestSignChanges(t,lastsubthresholdsample, WindowSizeInt);
|
|
if(mUseDirectionChangesLow || mUseDirectionChangesHigh)
|
|
dc = TestDirectionChanges(t,lastsubthresholdsample,WindowSizeInt);
|
|
|
|
//Now, go through the sound again, sample by sample.
|
|
for(i=samplesleft-1; i>WindowSizeInt; i--){
|
|
int tests = 0;
|
|
int testThreshold = 0;
|
|
//Update the test statistics
|
|
if(mUseEnergy)
|
|
{
|
|
TestEnergyUpdate(erg, WindowSizeInt,buffer[i],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(erg>mThresholdEnergy);
|
|
testThreshold++;
|
|
}
|
|
if(mUseSignChangesLow)
|
|
{
|
|
TestSignChangesUpdate(sc,WindowSizeInt,buffer[i],buffer[i+1],buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(sc < mThresholdSignChangesLower);
|
|
testThreshold++;
|
|
}
|
|
if(mUseSignChangesHigh)
|
|
{
|
|
TestSignChangesUpdate(sc,WindowSizeInt,buffer[i],buffer[i+1],buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(sc > mThresholdSignChangesUpper);
|
|
testThreshold++;
|
|
}
|
|
if(mUseDirectionChangesLow)
|
|
{
|
|
TestDirectionChangesUpdate(dc,WindowSizeInt,atrend,buffer[i],buffer[i+1],ztrend,buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(dc < mThresholdDirectionChangesLower);
|
|
testThreshold++;
|
|
}
|
|
if(mUseDirectionChangesHigh)
|
|
{
|
|
TestDirectionChangesUpdate(dc,WindowSizeInt,atrend,buffer[i],buffer[i+1],ztrend,buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(dc > mThresholdDirectionChangesUpper);
|
|
testThreshold++;
|
|
}
|
|
|
|
if(tests >= testThreshold)
|
|
{ //Finish off on the first hit
|
|
break;
|
|
}
|
|
}
|
|
|
|
//When we get here, i+lastsubthresholdsample is the best guess for where the word starts
|
|
delete [] buffer;
|
|
return lastsubthresholdsample - samplesleft + i;
|
|
}
|
|
else{
|
|
//If we failed to find anything, return the start position
|
|
return end ;
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
//Move froward from the start to find an OFF region.
|
|
sampleCount VoiceKey::OffForward (WaveTrack & t, sampleCount start, sampleCount len) {
|
|
|
|
if((mWindowSize) >= len+10){
|
|
wxMessageBox(_("Selection is too small to use voice key."));
|
|
|
|
return start;
|
|
}
|
|
else{
|
|
|
|
sampleCount lastsubthresholdsample; // keeps track of the sample number of the last sample to not exceed the threshold
|
|
|
|
|
|
//Change the millisecond-based parameters into sample-based parameters
|
|
double rate = t.GetRate(); //Translates seconds to samples
|
|
unsigned int WindowSizeInt = (unsigned int)(rate * mWindowSize); //Size of window to examine
|
|
unsigned int SilentWindowSizeInt = (unsigned int)(rate * mSilentWindowSize); //This much signal is necessary to trip key
|
|
|
|
int samplesleft = len - WindowSizeInt; //Indexes the number of samples remaining in the selection
|
|
lastsubthresholdsample = start; //start this off at the selection start
|
|
unsigned int i; //iterates through waveblock
|
|
int blockruns=0; //keeps track of the number of consecutive above-threshold blocks
|
|
int blocksize; //The final block may be smaller than WindowSizeInt, so use this
|
|
|
|
//This loop goes through the selection a block at a time. If a long enough run
|
|
//of above-threshold blocks occur, we return to the last sub-threshold block and
|
|
//go through one sample at a time.
|
|
//If there are fewer than 10 samples leftover, don't bother.
|
|
for(i = start; samplesleft >=10; i+=(WindowSizeInt-1) , samplesleft -= (WindowSizeInt -1)){
|
|
|
|
//Set blocksize so that it is the right size
|
|
if(samplesleft < (int)WindowSizeInt){
|
|
blocksize = samplesleft;
|
|
}
|
|
else{
|
|
blocksize = WindowSizeInt;
|
|
}
|
|
|
|
if(!AboveThreshold(t,i,blocksize))
|
|
{
|
|
blockruns++; //Hit
|
|
}
|
|
else
|
|
{
|
|
blockruns=0; //Above threshold--start over
|
|
lastsubthresholdsample = i;
|
|
}
|
|
|
|
//If the blockrun is long enough, break out of the loop early:
|
|
if(blockruns > mSilentWindowSize/mWindowSize)
|
|
break;
|
|
|
|
}
|
|
|
|
//Now, if we broke out early (samplesleft > 10), go back to the lastsubthresholdsample and look more carefully
|
|
if(samplesleft > 10){
|
|
|
|
|
|
//Calculate how many to scan through--we only have to go through (at most)
|
|
//the first window + 1 samples--but we need another window samples to draw from.
|
|
samplesleft = 2*WindowSizeInt+1;
|
|
|
|
//To speed things up, create a local buffer to store things in, to avoid the costly t.Get();
|
|
//Only go through the first SilentWindowSizeInt samples, and choose the first that trips the key.
|
|
sampleFormat *buffer = new sampleFormat[samplesleft];
|
|
t.Get((samplePtr)buffer, floatSample, lastsubthresholdsample,samplesleft);
|
|
|
|
//Initialize these trend markers atrend and ztrend. They keep track of the
|
|
//up/down trends at the start and end of the evaluation window.
|
|
int atrend = sgn(buffer[1]-buffer[0]);
|
|
int ztrend = sgn(buffer[WindowSizeInt+1]-buffer[WindowSizeInt]);
|
|
|
|
|
|
double erg=0;
|
|
double sc=0;
|
|
double dc=0;
|
|
|
|
//Get initial test statistic values.
|
|
if(mUseEnergy)
|
|
erg = TestEnergy(t, lastsubthresholdsample, WindowSizeInt);
|
|
if(mUseSignChangesLow || mUseSignChangesHigh)
|
|
sc = TestSignChanges(t,lastsubthresholdsample, WindowSizeInt);
|
|
if(mUseDirectionChangesLow || mUseDirectionChangesHigh)
|
|
dc = TestDirectionChanges(t,lastsubthresholdsample,WindowSizeInt);
|
|
|
|
//Now, go through the sound again, sample by sample.
|
|
for(i=0; i<SilentWindowSizeInt-WindowSizeInt;i++){
|
|
int tests = 0;
|
|
int testThreshold = 0;
|
|
//Update the test statistics
|
|
if(mUseEnergy)
|
|
{
|
|
TestEnergyUpdate(erg, WindowSizeInt,buffer[i],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(erg>mThresholdEnergy);
|
|
testThreshold++;
|
|
}
|
|
if(mUseSignChangesLow)
|
|
{
|
|
TestSignChangesUpdate(sc,WindowSizeInt,buffer[i],buffer[i+1],buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(sc < mThresholdSignChangesLower);
|
|
testThreshold++;
|
|
}
|
|
if(mUseSignChangesHigh)
|
|
{
|
|
TestSignChangesUpdate(sc,WindowSizeInt,buffer[i],buffer[i+1],buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(sc > mThresholdSignChangesUpper);
|
|
testThreshold++;
|
|
}
|
|
if(mUseDirectionChangesLow)
|
|
{
|
|
TestDirectionChangesUpdate(dc,WindowSizeInt,atrend,buffer[i],buffer[i+1],ztrend,buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(dc < mThresholdDirectionChangesLower);
|
|
testThreshold++;
|
|
}
|
|
if(mUseDirectionChangesHigh)
|
|
{
|
|
TestDirectionChangesUpdate(dc,WindowSizeInt,atrend,buffer[i],buffer[i+1],ztrend,buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(dc > mThresholdDirectionChangesUpper);
|
|
testThreshold++;
|
|
}
|
|
|
|
if(tests < testThreshold)
|
|
{ //Finish off on the first below-threshold block
|
|
break;
|
|
}
|
|
}
|
|
|
|
//When we get here, i+lastsubthresholdsample is the best guess for where the word starts
|
|
delete [] buffer;
|
|
return i+lastsubthresholdsample;
|
|
}
|
|
else{
|
|
//If we failed to find anything, return the start position
|
|
return start ;
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
//Move backward from the end to find an OFF region
|
|
sampleCount VoiceKey::OffBackward (WaveTrack & t, sampleCount end, sampleCount len) {
|
|
|
|
|
|
if((mWindowSize) >= len+10){
|
|
|
|
wxMessageBox(_("Selection is too small to use voice key."));
|
|
return end;
|
|
}
|
|
else{
|
|
|
|
sampleCount lastsubthresholdsample; // keeps track of the sample number of the last sample to not exceed the threshold
|
|
|
|
//Change the millisecond-based parameters into sample-based parameters
|
|
double rate = t.GetRate(); //Translates seconds to samples
|
|
unsigned int WindowSizeInt = (unsigned int)(rate * mWindowSize); //Size of window to examine
|
|
//unsigned int SilentWindowSizeInt = (unsigned int)(rate * mSilentWindowSize); //This much signal is necessary to trip key
|
|
|
|
int samplesleft = len - WindowSizeInt; //Indexes the number of samples remaining in the selection
|
|
lastsubthresholdsample = end; //start this off at the end
|
|
unsigned int i; //iterates through waveblock
|
|
int blockruns=0; //keeps track of the number of consecutive above-threshold blocks
|
|
int blocksize; //The final block may be smaller than WindowSizeInt, so use this
|
|
|
|
//This loop goes through the selection a block at a time in reverse order. If a long enough run
|
|
//of above-threshold blocks occur, we return to the last sub-threshold block and
|
|
//go through one sample at a time.
|
|
//If there are fewer than 10 samples leftover, don't bother.
|
|
for(i = end - WindowSizeInt; samplesleft >=10; i-=(WindowSizeInt-1) , samplesleft -= (WindowSizeInt -1)){
|
|
|
|
//Set blocksize so that it is the right size
|
|
if(samplesleft < (int)WindowSizeInt){
|
|
blocksize = samplesleft;
|
|
}
|
|
else{
|
|
blocksize = WindowSizeInt;
|
|
}
|
|
|
|
if(!AboveThreshold(t,i,blocksize))
|
|
{
|
|
|
|
blockruns++; //Hit
|
|
}
|
|
else
|
|
{
|
|
blockruns=0; //Miss--start over
|
|
lastsubthresholdsample = i+WindowSizeInt;
|
|
|
|
}
|
|
|
|
//If the blockrun is long enough, break out of the loop early:
|
|
if(blockruns > mSilentWindowSize/mWindowSize)
|
|
break;
|
|
|
|
}
|
|
|
|
//Now, if we broke out early (samplesleft > 10), go back to the lastsubthresholdsample and look more carefully
|
|
if(samplesleft > 10){
|
|
|
|
//Calculate how many to scan through--we only have to go through (at most)
|
|
//the first window + 1 samples--but we need another window samples to draw from.
|
|
samplesleft = 2*WindowSizeInt+1;
|
|
|
|
//To speed things up, create a local buffer to store things in, to avoid the costly t.Get();
|
|
//Only go through the first SilentWindowSizeInt samples, and choose the first that trips the key.
|
|
sampleFormat *buffer = new sampleFormat[samplesleft];
|
|
t.Get((samplePtr)buffer, floatSample, lastsubthresholdsample-samplesleft,samplesleft);
|
|
|
|
//Initialize these trend markers atrend and ztrend. They keep track of the
|
|
//up/down trends at the start and end of the evaluation window.
|
|
int atrend = sgn(buffer[samplesleft - 2]-buffer[samplesleft - 1]);
|
|
int ztrend = sgn(buffer[samplesleft - WindowSizeInt-2]-buffer[samplesleft - WindowSizeInt-2]);
|
|
|
|
double erg=0;
|
|
double sc=0;
|
|
double dc=0;
|
|
//Get initial test statistic values.
|
|
if(mUseEnergy)
|
|
erg = TestEnergy(t, lastsubthresholdsample, WindowSizeInt);
|
|
if(mUseSignChangesLow || mUseSignChangesHigh)
|
|
sc = TestSignChanges(t,lastsubthresholdsample, WindowSizeInt);
|
|
if(mUseDirectionChangesLow || mUseDirectionChangesHigh)
|
|
dc = TestDirectionChanges(t,lastsubthresholdsample,WindowSizeInt);
|
|
|
|
//Now, go through the sound again, sample by sample.
|
|
for(i=samplesleft-1; i>WindowSizeInt; i--){
|
|
|
|
int tests = 0;
|
|
int testThreshold = 0;
|
|
//Update the test statistics
|
|
if(mUseEnergy)
|
|
{
|
|
TestEnergyUpdate(erg, WindowSizeInt,buffer[i],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(erg>mThresholdEnergy);
|
|
testThreshold++;
|
|
}
|
|
if(mUseSignChangesLow)
|
|
{
|
|
TestSignChangesUpdate(sc,WindowSizeInt,buffer[i],buffer[i+1],buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(sc < mThresholdSignChangesLower);
|
|
testThreshold++;
|
|
}
|
|
if(mUseSignChangesHigh)
|
|
{
|
|
TestSignChangesUpdate(sc,WindowSizeInt,buffer[i],buffer[i+1],buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(sc > mThresholdSignChangesUpper);
|
|
testThreshold++;
|
|
}
|
|
if(mUseDirectionChangesLow)
|
|
{
|
|
TestDirectionChangesUpdate(dc,WindowSizeInt,atrend,buffer[i],buffer[i+1],ztrend,buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(dc < mThresholdDirectionChangesLower);
|
|
testThreshold++;
|
|
}
|
|
if(mUseDirectionChangesHigh)
|
|
{
|
|
TestDirectionChangesUpdate(dc,WindowSizeInt,atrend,buffer[i],buffer[i+1],ztrend,buffer[i+WindowSizeInt],buffer[i+WindowSizeInt+1]);
|
|
tests += (int)(dc > mThresholdDirectionChangesUpper);
|
|
testThreshold++;
|
|
}
|
|
|
|
|
|
|
|
if(tests < testThreshold)
|
|
{ //Finish off on the first hit
|
|
break;
|
|
}
|
|
}
|
|
|
|
//When we get here, i+lastsubthresholdsample is the best guess for where the word starts
|
|
delete [] buffer;
|
|
return lastsubthresholdsample - samplesleft + i;
|
|
}
|
|
else{
|
|
//If we failed to find anything, return the start position
|
|
return end ;
|
|
}
|
|
}
|
|
}
|
|
|
|
//This tests whether a specified block region is above or below threshold.
|
|
bool VoiceKey::AboveThreshold(WaveTrack & t, sampleCount start, sampleCount len)
|
|
{
|
|
|
|
double erg=0;
|
|
double sc=0;
|
|
double dc=0; //These store three statistics: energy, signchanges, and directionchanges
|
|
int tests =0; //Keeps track of how many statistics surpass the threshold.
|
|
int testThreshold=0; //Keeps track of the threshold.
|
|
|
|
//Calculate the test statistics
|
|
if(mUseEnergy)
|
|
{
|
|
testThreshold++;
|
|
erg = TestEnergy(t, start,len);
|
|
tests +=(int)(erg > mThresholdEnergy);
|
|
#if 0
|
|
std::cout << "Energy: " << erg << " " <<mThresholdEnergy << std::endl;
|
|
#endif
|
|
}
|
|
|
|
if(mUseSignChangesLow)
|
|
{
|
|
testThreshold++;
|
|
sc = TestSignChanges(t,start,len);
|
|
tests += (int)(sc < mThresholdSignChangesLower);
|
|
#if 0
|
|
std::cout << "SignChanges: " << sc << " " <<mThresholdSignChangesLower<< " < " << mThresholdSignChangesUpper << std::endl;
|
|
#endif
|
|
|
|
}
|
|
if(mUseSignChangesHigh)
|
|
{
|
|
testThreshold++;
|
|
sc = TestSignChanges(t,start,len);
|
|
tests += (int)(sc > mThresholdSignChangesUpper);
|
|
#if 0
|
|
std::cout << "SignChanges: " << sc << " " <<mThresholdSignChangesLower<< " < " << mThresholdSignChangesUpper << std::endl;
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
if(mUseDirectionChangesLow)
|
|
{
|
|
testThreshold++;
|
|
dc = TestDirectionChanges(t,start,len);
|
|
tests += (int)(dc < mThresholdDirectionChangesLower);
|
|
#if 0
|
|
std::cout << "DirectionChanges: " << dc << " " <<mThresholdDirectionChangesLower<< " < " << mThresholdDirectionChangesUpper << std::endl;
|
|
#endif
|
|
}
|
|
if(mUseDirectionChangesHigh)
|
|
{
|
|
testThreshold++;
|
|
dc = TestDirectionChanges(t,start,len);
|
|
tests += (int)(dc > mThresholdDirectionChangesUpper);
|
|
#if 0
|
|
std::cout << "DirectionChanges: " << dc << " " <<mThresholdDirectionChangesLower<< " < " << mThresholdDirectionChangesUpper << std::endl;
|
|
#endif
|
|
}
|
|
|
|
//Test whether we are above threshold (the number of stats)
|
|
return (tests >= testThreshold);
|
|
|
|
}
|
|
|
|
//This adjusts the threshold. Larger values of t expand the noise region,
|
|
//making more things be classified as noise (and requiring a stronger signal).
|
|
void VoiceKey::AdjustThreshold(double t){
|
|
|
|
mThresholdAdjustment = t;
|
|
mThresholdEnergy = mEnergyMean + mEnergySD * t;
|
|
mThresholdSignChangesUpper = mSignChangesMean + mSignChangesSD * t;
|
|
mThresholdSignChangesLower = mSignChangesMean - mSignChangesSD * t;
|
|
mThresholdDirectionChangesUpper = mDirectionChangesMean + mDirectionChangesSD * t;
|
|
mThresholdDirectionChangesLower = mDirectionChangesMean - mDirectionChangesSD * t;
|
|
};
|
|
|
|
|
|
//This 'calibrates' the voicekey to noise
|
|
void VoiceKey::CalibrateNoise(WaveTrack & t, sampleCount start, sampleCount len){
|
|
//To calibrate the noise, we need to scan the sample block just like in the voicekey and
|
|
//calculate the mean and standard deviation of the test statistics.
|
|
//Then, we set the BaselineThreshold to be one
|
|
|
|
wxBusyCursor busy;
|
|
|
|
//initialize some sample statistics: sums of X and X^2
|
|
|
|
double sumerg, sumerg2;
|
|
double sumsc, sumsc2;
|
|
double sumdc, sumdc2;
|
|
double erg, sc, dc;
|
|
// sampleFormat a1=0;
|
|
// sampleFormat a2=0;
|
|
// sampleFormat z1=0;
|
|
// sampleFormat z2=0; // keeps track of initial and final samples of window to enable updating formulae
|
|
int atrend=0; // equals sgn(a2-a1); Keeps track of trend at start of window
|
|
int ztrend=0; // equals sgn(z2-z1); Keeps track of trend at end of window
|
|
|
|
//Now, change the millisecond-based parameters into sample-based parameters
|
|
//(This depends on WaveTrack t)
|
|
double rate = t.GetRate();
|
|
unsigned int WindowSizeInt = (unsigned int)(rate * mWindowSize);
|
|
// unsigned int SignalWindowSizeInt = (unsigned int)(rate * mSignalWindowSize);
|
|
|
|
|
|
//Get the first test statistics
|
|
|
|
//Calibrate all of the statistic, because they might be
|
|
//changed later.
|
|
|
|
// if(mUseEnergy)
|
|
erg = TestEnergy(t, start, WindowSizeInt);
|
|
|
|
// if(mUseSignChanges)
|
|
sc = TestSignChanges(t,start, WindowSizeInt);
|
|
|
|
// if(mUseDirectionChanges)
|
|
dc = TestDirectionChanges(t,start,WindowSizeInt);
|
|
|
|
//Calculate initial values for the trend trackers--------------------//
|
|
//Get((samplePtr)a1, floatSample, start, 1);
|
|
// t.Get((samplePtr)a2, floatSample, start+1, 1);
|
|
// t.Get((samplePtr)z1, floatSample, start+WindowSizeInt-1, 1);
|
|
// t.Get((samplePtr)z2, floatSample, start+WindowSizeInt, 1);
|
|
atrend = -1;//sgn(a2-a1);
|
|
ztrend = -1;// sgn(z2-z1);
|
|
//-------------------------------------------------------------------//
|
|
|
|
|
|
sumerg =0.0;
|
|
sumerg2 = 0.0;
|
|
sumsc =0.0;
|
|
sumsc2 = 0.0;
|
|
sumdc =0.0;
|
|
sumdc2 =0.0;
|
|
|
|
|
|
// int n = len - WindowSizeInt; //This is how many samples we have
|
|
int samplesleft = len - WindowSizeInt;
|
|
int i;
|
|
int blocksize;
|
|
int samples=0;
|
|
|
|
for(i = start; samplesleft >=10 ; i += (WindowSizeInt -1), samplesleft -= (WindowSizeInt -1) )
|
|
{
|
|
//Take samples chunk-by-chunk.
|
|
//Normally, this should be in WindowSizeInt chunks, but at the end (if there are more than 10
|
|
//samples left) take a chunk that eats the rest of the samples.
|
|
|
|
samples++; //Increment the number of samples we have
|
|
if(samplesleft < (int)WindowSizeInt)
|
|
{
|
|
blocksize = samplesleft;
|
|
}
|
|
else
|
|
{
|
|
blocksize = WindowSizeInt;
|
|
}
|
|
|
|
erg = TestEnergy(t, i, blocksize);
|
|
sumerg +=(double)erg;
|
|
sumerg2 += pow((double)erg,2);
|
|
|
|
sc = TestSignChanges(t,i, blocksize);
|
|
sumsc += (double)sc;
|
|
sumsc2 += pow((double)sc,2);
|
|
|
|
|
|
dc = TestDirectionChanges(t,i,blocksize);
|
|
sumdc += (double)dc;
|
|
sumdc2 += pow((double)dc,2);
|
|
|
|
|
|
|
|
}
|
|
|
|
mEnergyMean = sumerg / samples;
|
|
mEnergySD = sqrt(sumerg2/samples - mEnergyMean*mEnergyMean);
|
|
|
|
mSignChangesMean = sumsc / samples;
|
|
mSignChangesSD = sqrt(sumsc2 / samples - mSignChangesMean * mSignChangesMean);
|
|
|
|
mDirectionChangesMean = sumdc / samples;
|
|
mDirectionChangesSD =sqrt(sumdc2 / samples - mDirectionChangesMean * mDirectionChangesMean) ;
|
|
|
|
|
|
|
|
|
|
|
|
wxString text = wxString::Format(_("Calibration Results\n"));
|
|
text += wxString::Format(_("Energy -- mean: %1.4f sd: (%1.4f)\n"),mEnergyMean,mEnergySD);
|
|
text+= wxString::Format(_("Sign Changes -- mean: %1.4f sd: (%1.4f)\n"),mSignChangesMean,mSignChangesSD);
|
|
text+= wxString::Format(_("Direction Changes -- mean: %1.4f sd: (%1.4f)\n"),mDirectionChangesMean,mDirectionChangesSD);
|
|
wxMessageDialog * stats = new wxMessageDialog(NULL, text,
|
|
wxT("Calibration Complete"),
|
|
wxOK | wxICON_INFORMATION,
|
|
wxPoint(-1,-1));
|
|
stats->ShowModal();
|
|
delete stats;
|
|
|
|
AdjustThreshold(mThresholdAdjustment);
|
|
}
|
|
|
|
|
|
void VoiceKey::SetKeyType(bool erg, bool scLow , bool scHigh,
|
|
bool dcLow, bool dcHigh)
|
|
{
|
|
mUseEnergy = erg;
|
|
mUseSignChangesLow = scLow;
|
|
mUseSignChangesHigh = scHigh;
|
|
mUseDirectionChangesLow = dcLow;
|
|
mUseDirectionChangesHigh = dcHigh;
|
|
}
|
|
|
|
|
|
//This might continue over a number of blocks.
|
|
double VoiceKey::TestEnergy (WaveTrack & t, sampleCount start, sampleCount len)
|
|
{
|
|
|
|
double sum = 1;
|
|
sampleCount s = start; //Keep track of start
|
|
sampleCount originalLen = len; //Keep track of the length of block to process (its not the length of t)
|
|
sampleCount blockSize = t.GetMaxBlockSize(); //Determine size of sampling buffer
|
|
|
|
|
|
if( blockSize > len)
|
|
blockSize = len;
|
|
float *buffer = new float[blockSize]; //Get a sampling buffer
|
|
|
|
while(len > 0)
|
|
{
|
|
sampleCount block = t.GetBestBlockSize(s); //Figure out how much to grab
|
|
if(block > len) block = len; //Don't grab too much!
|
|
|
|
t.Get((samplePtr)buffer,floatSample, s,block); //grab the block;
|
|
|
|
//Now, go through the block and calculate energy
|
|
for(int i = 0; i< block; i++)
|
|
{
|
|
sum += buffer[i]*buffer[i];
|
|
}
|
|
|
|
len -= block;
|
|
s += block;
|
|
}
|
|
|
|
delete [] buffer;
|
|
return sum / originalLen;
|
|
}
|
|
|
|
|
|
//This will update RMSE by adding one element and subtracting another
|
|
void VoiceKey::TestEnergyUpdate (double & prevErg, int len, const sampleFormat & drop, const sampleFormat & add)
|
|
{
|
|
//This is an updating formula for RMSE. It will only recalculate what's changed.
|
|
prevErg = prevErg + (double)(abs(add) - abs(drop))/len;
|
|
|
|
}
|
|
|
|
|
|
double VoiceKey::TestSignChanges(WaveTrack & t, sampleCount start, sampleCount len)
|
|
{
|
|
|
|
|
|
sampleCount s = start; //Keep track of start
|
|
sampleCount originalLen = len; //Keep track of the length of block to process (its not the length of t)
|
|
sampleCount blockSize = t.GetMaxBlockSize(); //Determine size of sampling buffer
|
|
unsigned long signchanges = 1;
|
|
int currentsign=0;
|
|
|
|
|
|
if( blockSize > len)
|
|
blockSize = len;
|
|
sampleFormat *buffer = new sampleFormat[blockSize]; //Get a sampling buffer
|
|
|
|
while(len > 0) {
|
|
|
|
sampleCount block = t.GetBestBlockSize(s); //Figure out how much to grab
|
|
if(block > len) block = len; //Don't grab too much!
|
|
|
|
t.Get((samplePtr)buffer,floatSample, s,block); //grab the block;
|
|
|
|
if (len == originalLen)
|
|
{
|
|
//The first time through, set stuff up special.
|
|
currentsign = sgn(buffer[0]);
|
|
}
|
|
|
|
//Now, go through the block and calculate zero crossings
|
|
|
|
for(int i = 0; i< block; i++)
|
|
{
|
|
if( sgn(buffer[i]) != currentsign)
|
|
{
|
|
currentsign = sgn(buffer[i]);
|
|
signchanges++;
|
|
}
|
|
|
|
}
|
|
len -= block;
|
|
s += block;
|
|
}
|
|
delete [] buffer;
|
|
return (double)signchanges/originalLen;
|
|
}
|
|
|
|
void VoiceKey::TestSignChangesUpdate(double & currentsignchanges, int len,
|
|
const sampleFormat & a1,
|
|
const sampleFormat & a2,
|
|
const sampleFormat & z1,
|
|
const sampleFormat & z2)
|
|
{
|
|
|
|
if(sgn(a1)!=sgn(a2)) currentsignchanges -= 1.0/len;
|
|
if(sgn(z1)!=sgn(z2)) currentsignchanges += 1.0/len;
|
|
|
|
}
|
|
|
|
|
|
double VoiceKey::TestDirectionChanges(WaveTrack & t, sampleCount start, sampleCount len)
|
|
{
|
|
|
|
|
|
sampleCount s = start; //Keep track of start
|
|
sampleCount originalLen = len; //Keep track of the length of block to process (its not the length of t)
|
|
sampleCount blockSize = t.GetMaxBlockSize(); //Determine size of sampling buffer
|
|
unsigned long directionchanges = 1;
|
|
sampleFormat lastval=sampleFormat(0);
|
|
int lastdirection=1;
|
|
|
|
if( blockSize > len)
|
|
blockSize = len;
|
|
sampleFormat *buffer = new sampleFormat[blockSize]; //Get a sampling buffer
|
|
|
|
while(len > 0) {
|
|
|
|
sampleCount block = t.GetBestBlockSize(s); //Figure out how much to grab
|
|
if(block > len) block = len; //Don't grab too much!
|
|
|
|
t.Get((samplePtr)buffer,floatSample, s,block); //grab the block;
|
|
|
|
if (len == originalLen) {
|
|
//The first time through, set stuff up special.
|
|
lastval = buffer[0];
|
|
}
|
|
|
|
//Now, go through the block and calculate zero crossings
|
|
|
|
|
|
for(int i = 0; i< block; i++){
|
|
|
|
if( sgn(buffer[i]-lastval) != lastdirection) {
|
|
directionchanges++;
|
|
lastdirection = sgn(buffer[i] - lastval);
|
|
}
|
|
lastval = buffer[i];
|
|
|
|
}
|
|
len -= block;
|
|
s += block;
|
|
}
|
|
delete [] buffer;
|
|
return (double)directionchanges/originalLen;
|
|
}
|
|
|
|
|
|
|
|
|
|
// This method does an updating by looking at the trends
|
|
// This will change currentdirections and atrend/trend, so be warned.
|
|
void VoiceKey::TestDirectionChangesUpdate(double & currentdirectionchanges, int len,
|
|
int & atrend, const sampleFormat & a1, const sampleFormat & a2,
|
|
int & ztrend, const sampleFormat & z1, const sampleFormat & z2)
|
|
{
|
|
|
|
if(sgn(a2 - a1)!= atrend ) {
|
|
//Here, the direction shifted for the item we're dropping.
|
|
currentdirectionchanges -= 1.0/len;
|
|
atrend = sgn(a2-a1);
|
|
}
|
|
if(sgn(z2 - z1)!= ztrend){
|
|
//Here, the direction shifts when we add an item
|
|
currentdirectionchanges += 1.0/len;
|
|
ztrend = sgn(z2-z1);
|
|
}
|
|
|
|
}
|
|
|
|
|
|
// Indentation settings for Vim and Emacs and unique identifier for Arch, a
|
|
// version control system. Please do not modify past this point.
|
|
//
|
|
// Local Variables:
|
|
// c-basic-offset: 3
|
|
// indent-tabs-mode: nil
|
|
// End:
|
|
//
|
|
// vim: et sts=3 sw=3
|
|
// arch-tag:
|
|
|