• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
3  *
4  *  Use of this source code is governed by a BSD-style license
5  *  that can be found in the LICENSE file in the root of the source
6  *  tree. An additional intellectual property rights grant can be found
7  *  in the file PATENTS.  All contributing project authors may
8  *  be found in the AUTHORS file in the root of the source tree.
9  */
10 
11 #include "webrtc/common_audio/vad/vad_gmm.h"
12 
13 #include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
14 #include "webrtc/typedefs.h"
15 
16 static const int32_t kCompVar = 22005;
17 static const int16_t kLog2Exp = 5909;  // log2(exp(1)) in Q12.
18 
19 // For a normal distribution, the probability of |input| is calculated and
20 // returned (in Q20). The formula for normal distributed probability is
21 //
22 // 1 / s * exp(-(x - m)^2 / (2 * s^2))
23 //
24 // where the parameters are given in the following Q domains:
25 // m = |mean| (Q7)
26 // s = |std| (Q7)
27 // x = |input| (Q4)
28 // in addition to the probability we output |delta| (in Q11) used when updating
29 // the noise/speech model.
WebRtcVad_GaussianProbability(int16_t input,int16_t mean,int16_t std,int16_t * delta)30 int32_t WebRtcVad_GaussianProbability(int16_t input,
31                                       int16_t mean,
32                                       int16_t std,
33                                       int16_t* delta) {
34   int16_t tmp16, inv_std, inv_std2, exp_value = 0;
35   int32_t tmp32;
36 
37   // Calculate |inv_std| = 1 / s, in Q10.
38   // 131072 = 1 in Q17, and (|std| >> 1) is for rounding instead of truncation.
39   // Q-domain: Q17 / Q7 = Q10.
40   tmp32 = (int32_t) 131072 + (int32_t) (std >> 1);
41   inv_std = (int16_t) WebRtcSpl_DivW32W16(tmp32, std);
42 
43   // Calculate |inv_std2| = 1 / s^2, in Q14.
44   tmp16 = (inv_std >> 2);  // Q10 -> Q8.
45   // Q-domain: (Q8 * Q8) >> 2 = Q14.
46   inv_std2 = (int16_t)((tmp16 * tmp16) >> 2);
47   // TODO(bjornv): Investigate if changing to
48   // inv_std2 = (int16_t)((inv_std * inv_std) >> 6);
49   // gives better accuracy.
50 
51   tmp16 = (input << 3);  // Q4 -> Q7
52   tmp16 = tmp16 - mean;  // Q7 - Q7 = Q7
53 
54   // To be used later, when updating noise/speech model.
55   // |delta| = (x - m) / s^2, in Q11.
56   // Q-domain: (Q14 * Q7) >> 10 = Q11.
57   *delta = (int16_t)((inv_std2 * tmp16) >> 10);
58 
59   // Calculate the exponent |tmp32| = (x - m)^2 / (2 * s^2), in Q10. Replacing
60   // division by two with one shift.
61   // Q-domain: (Q11 * Q7) >> 8 = Q10.
62   tmp32 = (*delta * tmp16) >> 9;
63 
64   // If the exponent is small enough to give a non-zero probability we calculate
65   // |exp_value| ~= exp(-(x - m)^2 / (2 * s^2))
66   //             ~= exp2(-log2(exp(1)) * |tmp32|).
67   if (tmp32 < kCompVar) {
68     // Calculate |tmp16| = log2(exp(1)) * |tmp32|, in Q10.
69     // Q-domain: (Q12 * Q10) >> 12 = Q10.
70     tmp16 = (int16_t)((kLog2Exp * tmp32) >> 12);
71     tmp16 = -tmp16;
72     exp_value = (0x0400 | (tmp16 & 0x03FF));
73     tmp16 ^= 0xFFFF;
74     tmp16 >>= 10;
75     tmp16 += 1;
76     // Get |exp_value| = exp(-|tmp32|) in Q10.
77     exp_value >>= tmp16;
78   }
79 
80   // Calculate and return (1 / s) * exp(-(x - m)^2 / (2 * s^2)), in Q20.
81   // Q-domain: Q10 * Q10 = Q20.
82   return inv_std * exp_value;
83 }
84