generic-library/webrtc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Refactoring vad_gmm.[c/h].

Browse Source 
- Changed to stdint.
- Replaced SHIFT macros.
- Variable name changes.
- Style changes.
- Comments updates.
- Added a unit test.
Review URL: http://webrtc-codereview.appspot.com/323011

git-svn-id: http://webrtc.googlecode.com/svn/trunk@1249 4adac7df-926f-26a2-2b94-8c16560cd09d
This commit is contained in:
bjornv@webrtc.org 2011-12-20 14:08:34 +00:00
parent 42d07f0c58
commit c68f80a70a
3 changed files with 114 additions and 84 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

110
src/common_audio/vad/vad_gmm.c
View File

@ -8,68 +8,76 @@
* be found in the AUTHORS file in the root of the source tree. * be found in the AUTHORS file in the root of the source tree.
*/ */
/*
* This file includes the implementation of the internal VAD call
* WebRtcVad_GaussianProbability. For function description, see vad_gmm.h.
*/
#include "vad_gmm.h" #include "vad_gmm.h"
#include "signal_processing_library.h" #include "signal_processing_library.h"
#include "typedefs.h" #include "typedefs.h"
static const WebRtc_Word32 kCompVar = 22005; static const int32_t kCompVar = 22005;
// Constant log2(exp(1)) in Q12 static const int16_t kLog2Exp = 5909; // log2(exp(1)) in Q12.
static const WebRtc_Word16 kLog10Const = 5909;
WebRtc_Word32 WebRtcVad_GaussianProbability(WebRtc_Word16 in_sample, // For a normal distribution, the probability of |input| is calculated and
WebRtc_Word16 mean, // returned (in Q20). The formula for normal distributed probability is
WebRtc_Word16 std, //
WebRtc_Word16 *delta) // 1 / s * exp(-(x - m)^2 / (2 * s^2))
{ //
WebRtc_Word16 tmp16, tmpDiv, tmpDiv2, expVal, tmp16_1, tmp16_2; // where the parameters are given in the following Q domains:
WebRtc_Word32 tmp32, y32; // m = |mean| (Q7)
// s = |std| (Q7)
// x = |input| (Q4)
// in addition to the probability we output |delta| (in Q11) used when updating
// the noise/speech model.
int32_t WebRtcVad_GaussianProbability(int16_t input,
int16_t mean,
int16_t std,
int16_t* delta) {
int16_t tmp16, inv_std, inv_std2, exp_value = 0;
int32_t tmp32;
// Calculate tmpDiv=1/std, in Q10 // Calculate |inv_std| = 1 / s, in Q10.
tmp32 = (WebRtc_Word32)WEBRTC_SPL_RSHIFT_W16(std,1) + (WebRtc_Word32)131072; // 1 in Q17 // 131072 = 1 in Q17, and (|std| >> 1) is for rounding instead of truncation.
tmpDiv = (WebRtc_Word16)WebRtcSpl_DivW32W16(tmp32, std); // Q17/Q7 = Q10 // Q-domain: Q17 / Q7 = Q10.
tmp32 = (int32_t) 131072 + (int32_t) (std >> 1);
inv_std = (int16_t) WebRtcSpl_DivW32W16(tmp32, std);
// Calculate tmpDiv2=1/std^2, in Q14 // Calculate |inv_std2| = 1 / s^2, in Q14.
tmp16 = WEBRTC_SPL_RSHIFT_W16(tmpDiv, 2); // From Q10 to Q8 tmp16 = (inv_std >> 2); // Q10 -> Q8.
tmpDiv2 = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16_RSFT(tmp16, tmp16, 2); // (Q8 * Q8)>>2 = Q14 // Q-domain: (Q8 * Q8) >> 2 = Q14.
inv_std2 = (int16_t) WEBRTC_SPL_MUL_16_16_RSFT(tmp16, tmp16, 2);
// TODO(bjornv): Investigate if changing to
// |inv_std2| = (int16_t) WEBRTC_SPL_MUL_16_16_RSFT(|inv_std|, |inv_std|, 6);
// gives better accuracy.
tmp16 = WEBRTC_SPL_LSHIFT_W16(in_sample, 3); // Q7 tmp16 = (input << 3); // Q4 -> Q7
tmp16 = tmp16 - mean; // Q7 - Q7 = Q7 tmp16 = tmp16 - mean; // Q7 - Q7 = Q7
// To be used later, when updating noise/speech model // To be used later, when updating noise/speech model.
// delta = (x-m)/std^2, in Q11 // |delta| = (x - m) / s^2, in Q11.
*delta = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16_RSFT(tmpDiv2, tmp16, 10); //(Q14*Q7)>>10 = Q11 // Q-domain: (Q14 * Q7) >> 10 = Q11.
*delta = (int16_t) WEBRTC_SPL_MUL_16_16_RSFT(inv_std2, tmp16, 10);
// Calculate tmp32=(x-m)^2/(2*std^2), in Q10 // Calculate the exponent |tmp32| = (x - m)^2 / (2 * s^2), in Q10. Replacing
tmp32 = (WebRtc_Word32)WEBRTC_SPL_MUL_16_16_RSFT(*delta, tmp16, 9); // One shift for /2 // division by two with one shift.
// Q-domain: (Q11 * Q7) >> 8 = Q10.
tmp32 = WEBRTC_SPL_MUL_16_16_RSFT(*delta, tmp16, 9);
// Calculate expVal ~= exp(-(x-m)^2/(2*std^2)) ~= exp2(-log2(exp(1))*tmp32) // If the exponent is small enough to give a non-zero probability we calculate
if (tmp32 < kCompVar) // |exp_value| ~= exp(-(x - m)^2 / (2 * s^2))
{ // ~= exp2(-log2(exp(1)) * |tmp32|).
// Calculate tmp16 = log2(exp(1))*tmp32 , in Q10 if (tmp32 < kCompVar) {
tmp16 = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16_RSFT((WebRtc_Word16)tmp32, // Calculate |tmp16| = log2(exp(1)) * |tmp32|, in Q10.
kLog10Const, 12); // Q-domain: (Q12 * Q10) >> 12 = Q10.
tmp16 = -tmp16; tmp16 = (int16_t) WEBRTC_SPL_MUL_16_16_RSFT(kLog2Exp, (int16_t) tmp32, 12);
tmp16_2 = (WebRtc_Word16)(0x0400 | (tmp16 & 0x03FF)); tmp16 = -tmp16;
tmp16_1 = (WebRtc_Word16)(tmp16 ^ 0xFFFF); exp_value = (0x0400 | (tmp16 & 0x03FF));
tmp16 = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W16(tmp16_1, 10); tmp16 ^= 0xFFFF;
tmp16 += 1; tmp16 >>= 10;
// Calculate expVal=log2(-tmp32), in Q10 tmp16 += 1;
expVal = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((WebRtc_Word32)tmp16_2, tmp16); // Get |exp_value| = exp(-|tmp32|) in Q10.
exp_value >>= tmp16;
}
} else // Calculate and return (1 / s) * exp(-(x - m)^2 / (2 * s^2)), in Q20.
{ // Q-domain: Q10 * Q10 = Q20.
expVal = 0; return WEBRTC_SPL_MUL_16_16(inv_std, exp_value);
}
// Calculate y32=(1/std)*exp(-(x-m)^2/(2*std^2)), in Q20
y32 = WEBRTC_SPL_MUL_16_16(tmpDiv, expVal); // Q10 * Q10 = Q20
return y32; // Q20
} }

56
src/common_audio/vad/vad_gmm.h
View File

@ -8,40 +8,32 @@
* be found in the AUTHORS file in the root of the source tree. * be found in the AUTHORS file in the root of the source tree.
*/ */
// Gaussian probability calculations internally used in vad_core.c.
/* #ifndef WEBRTC_COMMON_AUDIO_VAD_VAD_GMM_H_
* This header file includes the description of the internal VAD call #define WEBRTC_COMMON_AUDIO_VAD_VAD_GMM_H_
* WebRtcVad_GaussianProbability.
*/
#ifndef WEBRTC_VAD_GMM_H_
#define WEBRTC_VAD_GMM_H_
#include "typedefs.h" #include "typedefs.h"
/**************************************************************************** // Calculates the probability for |input|, given that |input| comes from a
* WebRtcVad_GaussianProbability(...) // normal distribution with mean and standard deviation (|mean|, |std|).
* //
* This function calculates the probability for the value 'in_sample', given that in_sample // Inputs:
* comes from a normal distribution with mean 'mean' and standard deviation 'std'. // - input : input sample in Q4.
* // - mean : mean input in the statistical model, Q7.
* Input: // - std : standard deviation, Q7.
* - in_sample : Input sample in Q4 //
* - mean : mean value in the statistical model, Q7 // Output:
* - std : standard deviation, Q7 //
* // - delta : input used when updating the model, Q11.
* Output: // |delta| = (|input| - |mean|) / |std|^2.
* //
* - delta : Value used when updating the model, Q11 // Return:
* // (probability for |input|) =
* Return: // 1 / |std| * exp(-(|input| - |mean|)^2 / (2 * |std|^2));
* - out : out = 1/std * exp(-(x-m)^2/(2*std^2)); int32_t WebRtcVad_GaussianProbability(int16_t input,
* Probability for x. int16_t mean,
* int16_t std,
*/ int16_t* delta);
WebRtc_Word32 WebRtcVad_GaussianProbability(WebRtc_Word16 in_sample,
WebRtc_Word16 mean,
WebRtc_Word16 std,
WebRtc_Word16 *delta);
#endif // WEBRTC_VAD_GMM_H_ #endif // WEBRTC_COMMON_AUDIO_VAD_VAD_GMM_H_

32
src/common_audio/vad/vad_unittest.cc
View File

@ -8,13 +8,20 @@
* be found in the AUTHORS file in the root of the source tree. * be found in the AUTHORS file in the root of the source tree.
*/ */
#include <stddef.h> // size_t #include <stddef.h> // size_t
#include <stdlib.h> #include <stdlib.h>
#include "gtest/gtest.h" #include "gtest/gtest.h"
#include "typedefs.h" #include "typedefs.h"
#include "webrtc_vad.h" #include "webrtc_vad.h"
#ifdef __cplusplus
extern "C"
{
#include "vad_gmm.h"
}
#endif
namespace webrtc { namespace webrtc {
namespace { namespace {
const int16_t kModes[] = { 0, 1, 2, 3 }; const int16_t kModes[] = { 0, 1, 2, 3 };
@ -152,6 +159,29 @@ TEST_F(VadTest, ApiTest) {
EXPECT_EQ(0, WebRtcVad_Free(handle)); EXPECT_EQ(0, WebRtcVad_Free(handle));
} }
TEST_F(VadTest, GMMTests) {
int16_t delta = 0;
// Input value at mean.
EXPECT_EQ(1048576, WebRtcVad_GaussianProbability(0, 0, 128, &delta));
EXPECT_EQ(0, delta);
EXPECT_EQ(1048576, WebRtcVad_GaussianProbability(16, 128, 128, &delta));
EXPECT_EQ(0, delta);
EXPECT_EQ(1048576, WebRtcVad_GaussianProbability(-16, -128, 128, &delta));
EXPECT_EQ(0, delta);
// Largest possible input to give non-zero probability.
EXPECT_EQ(1024, WebRtcVad_GaussianProbability(59, 0, 128, &delta));
EXPECT_EQ(7552, delta);
EXPECT_EQ(1024, WebRtcVad_GaussianProbability(75, 128, 128, &delta));
EXPECT_EQ(7552, delta);
EXPECT_EQ(1024, WebRtcVad_GaussianProbability(-75, -128, 128, &delta));
EXPECT_EQ(-7552, delta);
// Too large input, should give zero probability.
EXPECT_EQ(0, WebRtcVad_GaussianProbability(105, 0, 128, &delta));
EXPECT_EQ(13440, delta);
}
// TODO(bjornv): Add a process test, run on file. // TODO(bjornv): Add a process test, run on file.
} // namespace } // namespace