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 "webrtc/common_audio/vad/vad_sp.h" 12 13 #include <assert.h> 14 15 #include "webrtc/common_audio/signal_processing/include/signal_processing_library.h" 16 #include "webrtc/common_audio/vad/vad_core.h" 17 #include "webrtc/typedefs.h" 18 19 // Allpass filter coefficients, upper and lower, in Q13. 20 // Upper: 0.64, Lower: 0.17. 21 static const int16_t kAllPassCoefsQ13[2] = { 5243, 1392 }; // Q13. 22 static const int16_t kSmoothingDown = 6553; // 0.2 in Q15. 23 static const int16_t kSmoothingUp = 32439; // 0.99 in Q15. 24 25 // TODO(bjornv): Move this function to vad_filterbank.c. 26 // Downsampling filter based on splitting filter and allpass functions. 27 void WebRtcVad_Downsampling(const int16_t* signal_in, 28 int16_t* signal_out, 29 int32_t* filter_state, 30 size_t in_length) { 31 int16_t tmp16_1 = 0, tmp16_2 = 0; 32 int32_t tmp32_1 = filter_state[0]; 33 int32_t tmp32_2 = filter_state[1]; 34 size_t n = 0; 35 // Downsampling by 2 gives half length. 36 size_t half_length = (in_length >> 1); 37 38 // Filter coefficients in Q13, filter state in Q0. 39 for (n = 0; n < half_length; n++) { 40 // All-pass filtering upper branch. 41 tmp16_1 = (int16_t) ((tmp32_1 >> 1) + 42 ((kAllPassCoefsQ13[0] * *signal_in) >> 14)); 43 *signal_out = tmp16_1; 44 tmp32_1 = (int32_t)(*signal_in++) - ((kAllPassCoefsQ13[0] * tmp16_1) >> 12); 45 46 // All-pass filtering lower branch. 47 tmp16_2 = (int16_t) ((tmp32_2 >> 1) + 48 ((kAllPassCoefsQ13[1] * *signal_in) >> 14)); 49 *signal_out++ += tmp16_2; 50 tmp32_2 = (int32_t)(*signal_in++) - ((kAllPassCoefsQ13[1] * tmp16_2) >> 12); 51 } 52 // Store the filter states. 53 filter_state[0] = tmp32_1; 54 filter_state[1] = tmp32_2; 55 } 56 57 // Inserts |feature_value| into |low_value_vector|, if it is one of the 16 58 // smallest values the last 100 frames. Then calculates and returns the median 59 // of the five smallest values. 60 int16_t WebRtcVad_FindMinimum(VadInstT* self, 61 int16_t feature_value, 62 int channel) { 63 int i = 0, j = 0; 64 int position = -1; 65 // Offset to beginning of the 16 minimum values in memory. 66 const int offset = (channel << 4); 67 int16_t current_median = 1600; 68 int16_t alpha = 0; 69 int32_t tmp32 = 0; 70 // Pointer to memory for the 16 minimum values and the age of each value of 71 // the |channel|. 72 int16_t* age = &self->index_vector[offset]; 73 int16_t* smallest_values = &self->low_value_vector[offset]; 74 75 assert(channel < kNumChannels); 76 77 // Each value in |smallest_values| is getting 1 loop older. Update |age|, and 78 // remove old values. 79 for (i = 0; i < 16; i++) { 80 if (age[i] != 100) { 81 age[i]++; 82 } else { 83 // Too old value. Remove from memory and shift larger values downwards. 84 for (j = i; j < 16; j++) { 85 smallest_values[j] = smallest_values[j + 1]; 86 age[j] = age[j + 1]; 87 } 88 age[15] = 101; 89 smallest_values[15] = 10000; 90 } 91 } 92 93 // Check if |feature_value| is smaller than any of the values in 94 // |smallest_values|. If so, find the |position| where to insert the new value 95 // (|feature_value|). 96 if (feature_value < smallest_values[7]) { 97 if (feature_value < smallest_values[3]) { 98 if (feature_value < smallest_values[1]) { 99 if (feature_value < smallest_values[0]) { 100 position = 0; 101 } else { 102 position = 1; 103 } 104 } else if (feature_value < smallest_values[2]) { 105 position = 2; 106 } else { 107 position = 3; 108 } 109 } else if (feature_value < smallest_values[5]) { 110 if (feature_value < smallest_values[4]) { 111 position = 4; 112 } else { 113 position = 5; 114 } 115 } else if (feature_value < smallest_values[6]) { 116 position = 6; 117 } else { 118 position = 7; 119 } 120 } else if (feature_value < smallest_values[15]) { 121 if (feature_value < smallest_values[11]) { 122 if (feature_value < smallest_values[9]) { 123 if (feature_value < smallest_values[8]) { 124 position = 8; 125 } else { 126 position = 9; 127 } 128 } else if (feature_value < smallest_values[10]) { 129 position = 10; 130 } else { 131 position = 11; 132 } 133 } else if (feature_value < smallest_values[13]) { 134 if (feature_value < smallest_values[12]) { 135 position = 12; 136 } else { 137 position = 13; 138 } 139 } else if (feature_value < smallest_values[14]) { 140 position = 14; 141 } else { 142 position = 15; 143 } 144 } 145 146 // If we have detected a new small value, insert it at the correct position 147 // and shift larger values up. 148 if (position > -1) { 149 for (i = 15; i > position; i--) { 150 smallest_values[i] = smallest_values[i - 1]; 151 age[i] = age[i - 1]; 152 } 153 smallest_values[position] = feature_value; 154 age[position] = 1; 155 } 156 157 // Get |current_median|. 158 if (self->frame_counter > 2) { 159 current_median = smallest_values[2]; 160 } else if (self->frame_counter > 0) { 161 current_median = smallest_values[0]; 162 } 163 164 // Smooth the median value. 165 if (self->frame_counter > 0) { 166 if (current_median < self->mean_value[channel]) { 167 alpha = kSmoothingDown; // 0.2 in Q15. 168 } else { 169 alpha = kSmoothingUp; // 0.99 in Q15. 170 } 171 } 172 tmp32 = (alpha + 1) * self->mean_value[channel]; 173 tmp32 += (WEBRTC_SPL_WORD16_MAX - alpha) * current_median; 174 tmp32 += 16384; 175 self->mean_value[channel] = (int16_t) (tmp32 >> 15); 176 177 return self->mean_value[channel]; 178 } 179