1 /*
2 * Copyright (c) 2012 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 "common_audio/vad/vad_sp.h"
12
13 #include "rtc_base/checks.h"
14 #include "common_audio/signal_processing/include/signal_processing_library.h"
15 #include "common_audio/vad/vad_core.h"
16
17 // Allpass filter coefficients, upper and lower, in Q13.
18 // Upper: 0.64, Lower: 0.17.
19 static const int16_t kAllPassCoefsQ13[2] = { 5243, 1392 }; // Q13.
20 static const int16_t kSmoothingDown = 6553; // 0.2 in Q15.
21 static const int16_t kSmoothingUp = 32439; // 0.99 in Q15.
22
23 // TODO(bjornv): Move this function to vad_filterbank.c.
24 // Downsampling filter based on splitting filter and allpass functions.
WebRtcVad_Downsampling(const int16_t * signal_in,int16_t * signal_out,int32_t * filter_state,size_t in_length)25 void WebRtcVad_Downsampling(const int16_t* signal_in,
26 int16_t* signal_out,
27 int32_t* filter_state,
28 size_t in_length) {
29 int16_t tmp16_1 = 0, tmp16_2 = 0;
30 int32_t tmp32_1 = filter_state[0];
31 int32_t tmp32_2 = filter_state[1];
32 size_t n = 0;
33 // Downsampling by 2 gives half length.
34 size_t half_length = (in_length >> 1);
35
36 // Filter coefficients in Q13, filter state in Q0.
37 for (n = 0; n < half_length; n++) {
38 // All-pass filtering upper branch.
39 tmp16_1 = (int16_t) ((tmp32_1 >> 1) +
40 ((kAllPassCoefsQ13[0] * *signal_in) >> 14));
41 *signal_out = tmp16_1;
42 tmp32_1 = (int32_t)(*signal_in++) - ((kAllPassCoefsQ13[0] * tmp16_1) >> 12);
43
44 // All-pass filtering lower branch.
45 tmp16_2 = (int16_t) ((tmp32_2 >> 1) +
46 ((kAllPassCoefsQ13[1] * *signal_in) >> 14));
47 *signal_out++ += tmp16_2;
48 tmp32_2 = (int32_t)(*signal_in++) - ((kAllPassCoefsQ13[1] * tmp16_2) >> 12);
49 }
50 // Store the filter states.
51 filter_state[0] = tmp32_1;
52 filter_state[1] = tmp32_2;
53 }
54
55 // Inserts |feature_value| into |low_value_vector|, if it is one of the 16
56 // smallest values the last 100 frames. Then calculates and returns the median
57 // of the five smallest values.
WebRtcVad_FindMinimum(VadInstT * self,int16_t feature_value,int channel)58 int16_t WebRtcVad_FindMinimum(VadInstT* self,
59 int16_t feature_value,
60 int channel) {
61 int i = 0, j = 0;
62 int position = -1;
63 // Offset to beginning of the 16 minimum values in memory.
64 const int offset = (channel << 4);
65 int16_t current_median = 1600;
66 int16_t alpha = 0;
67 int32_t tmp32 = 0;
68 // Pointer to memory for the 16 minimum values and the age of each value of
69 // the |channel|.
70 int16_t* age = &self->index_vector[offset];
71 int16_t* smallest_values = &self->low_value_vector[offset];
72
73 RTC_DCHECK_LT(channel, kNumChannels);
74
75 // Each value in |smallest_values| is getting 1 loop older. Update |age|, and
76 // remove old values.
77 for (i = 0; i < 16; i++) {
78 if (age[i] != 100) {
79 age[i]++;
80 } else {
81 // Too old value. Remove from memory and shift larger values downwards.
82 for (j = i; j < 15; j++) {
83 smallest_values[j] = smallest_values[j + 1];
84 age[j] = age[j + 1];
85 }
86 age[15] = 101;
87 smallest_values[15] = 10000;
88 }
89 }
90
91 // Check if |feature_value| is smaller than any of the values in
92 // |smallest_values|. If so, find the |position| where to insert the new value
93 // (|feature_value|).
94 if (feature_value < smallest_values[7]) {
95 if (feature_value < smallest_values[3]) {
96 if (feature_value < smallest_values[1]) {
97 if (feature_value < smallest_values[0]) {
98 position = 0;
99 } else {
100 position = 1;
101 }
102 } else if (feature_value < smallest_values[2]) {
103 position = 2;
104 } else {
105 position = 3;
106 }
107 } else if (feature_value < smallest_values[5]) {
108 if (feature_value < smallest_values[4]) {
109 position = 4;
110 } else {
111 position = 5;
112 }
113 } else if (feature_value < smallest_values[6]) {
114 position = 6;
115 } else {
116 position = 7;
117 }
118 } else if (feature_value < smallest_values[15]) {
119 if (feature_value < smallest_values[11]) {
120 if (feature_value < smallest_values[9]) {
121 if (feature_value < smallest_values[8]) {
122 position = 8;
123 } else {
124 position = 9;
125 }
126 } else if (feature_value < smallest_values[10]) {
127 position = 10;
128 } else {
129 position = 11;
130 }
131 } else if (feature_value < smallest_values[13]) {
132 if (feature_value < smallest_values[12]) {
133 position = 12;
134 } else {
135 position = 13;
136 }
137 } else if (feature_value < smallest_values[14]) {
138 position = 14;
139 } else {
140 position = 15;
141 }
142 }
143
144 // If we have detected a new small value, insert it at the correct position
145 // and shift larger values up.
146 if (position > -1) {
147 for (i = 15; i > position; i--) {
148 smallest_values[i] = smallest_values[i - 1];
149 age[i] = age[i - 1];
150 }
151 smallest_values[position] = feature_value;
152 age[position] = 1;
153 }
154
155 // Get |current_median|.
156 if (self->frame_counter > 2) {
157 current_median = smallest_values[2];
158 } else if (self->frame_counter > 0) {
159 current_median = smallest_values[0];
160 }
161
162 // Smooth the median value.
163 if (self->frame_counter > 0) {
164 if (current_median < self->mean_value[channel]) {
165 alpha = kSmoothingDown; // 0.2 in Q15.
166 } else {
167 alpha = kSmoothingUp; // 0.99 in Q15.
168 }
169 }
170 tmp32 = (alpha + 1) * self->mean_value[channel];
171 tmp32 += (WEBRTC_SPL_WORD16_MAX - alpha) * current_median;
172 tmp32 += 16384;
173 self->mean_value[channel] = (int16_t) (tmp32 >> 15);
174
175 return self->mean_value[channel];
176 }
177