• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Copyright (c) 2013 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/modules/audio_processing/transient/transient_detector.h"
12 
13 #include <assert.h>
14 #include <float.h>
15 #include <math.h>
16 #include <string.h>
17 
18 #include "webrtc/modules/audio_processing/transient/common.h"
19 #include "webrtc/modules/audio_processing/transient/daubechies_8_wavelet_coeffs.h"
20 #include "webrtc/modules/audio_processing/transient/moving_moments.h"
21 #include "webrtc/modules/audio_processing/transient/wpd_tree.h"
22 
23 namespace webrtc {
24 
25 static const int kTransientLengthMs = 30;
26 static const int kChunksAtStartupLeftToDelete =
27     kTransientLengthMs / ts::kChunkSizeMs;
28 static const float kDetectThreshold = 16.f;
29 
TransientDetector(int sample_rate_hz)30 TransientDetector::TransientDetector(int sample_rate_hz)
31     : samples_per_chunk_(sample_rate_hz * ts::kChunkSizeMs / 1000),
32       last_first_moment_(),
33       last_second_moment_(),
34       chunks_at_startup_left_to_delete_(kChunksAtStartupLeftToDelete),
35       reference_energy_(1.f),
36       using_reference_(false) {
37   assert(sample_rate_hz == ts::kSampleRate8kHz ||
38          sample_rate_hz == ts::kSampleRate16kHz ||
39          sample_rate_hz == ts::kSampleRate32kHz ||
40          sample_rate_hz == ts::kSampleRate48kHz);
41   int samples_per_transient = sample_rate_hz * kTransientLengthMs / 1000;
42   // Adjustment to avoid data loss while downsampling, making
43   // |samples_per_chunk_| and |samples_per_transient| always divisible by
44   // |kLeaves|.
45   samples_per_chunk_ -= samples_per_chunk_ % kLeaves;
46   samples_per_transient -= samples_per_transient % kLeaves;
47 
48   tree_leaves_data_length_ = samples_per_chunk_ / kLeaves;
49   wpd_tree_.reset(new WPDTree(samples_per_chunk_,
50                               kDaubechies8HighPassCoefficients,
51                               kDaubechies8LowPassCoefficients,
52                               kDaubechies8CoefficientsLength,
53                               kLevels));
54   for (size_t i = 0; i < kLeaves; ++i) {
55     moving_moments_[i].reset(
56         new MovingMoments(samples_per_transient / kLeaves));
57   }
58 
59   first_moments_.reset(new float[tree_leaves_data_length_]);
60   second_moments_.reset(new float[tree_leaves_data_length_]);
61 
62   for (int i = 0; i < kChunksAtStartupLeftToDelete; ++i) {
63     previous_results_.push_back(0.f);
64   }
65 }
66 
~TransientDetector()67 TransientDetector::~TransientDetector() {}
68 
Detect(const float * data,size_t data_length,const float * reference_data,size_t reference_length)69 float TransientDetector::Detect(const float* data,
70                                 size_t data_length,
71                                 const float* reference_data,
72                                 size_t reference_length) {
73   assert(data && data_length == samples_per_chunk_);
74 
75   // TODO(aluebs): Check if these errors can logically happen and if not assert
76   // on them.
77   if (wpd_tree_->Update(data, samples_per_chunk_) != 0) {
78     return -1.f;
79   }
80 
81   float result = 0.f;
82 
83   for (size_t i = 0; i < kLeaves; ++i) {
84     WPDNode* leaf = wpd_tree_->NodeAt(kLevels, i);
85 
86     moving_moments_[i]->CalculateMoments(leaf->data(),
87                                          tree_leaves_data_length_,
88                                          first_moments_.get(),
89                                          second_moments_.get());
90 
91     // Add value delayed (Use the last moments from the last call to Detect).
92     float unbiased_data = leaf->data()[0] - last_first_moment_[i];
93     result +=
94         unbiased_data * unbiased_data / (last_second_moment_[i] + FLT_MIN);
95 
96     // Add new values.
97     for (size_t j = 1; j < tree_leaves_data_length_; ++j) {
98       unbiased_data = leaf->data()[j] - first_moments_[j - 1];
99       result +=
100           unbiased_data * unbiased_data / (second_moments_[j - 1] + FLT_MIN);
101     }
102 
103     last_first_moment_[i] = first_moments_[tree_leaves_data_length_ - 1];
104     last_second_moment_[i] = second_moments_[tree_leaves_data_length_ - 1];
105   }
106 
107   result /= tree_leaves_data_length_;
108 
109   result *= ReferenceDetectionValue(reference_data, reference_length);
110 
111   if (chunks_at_startup_left_to_delete_ > 0) {
112     chunks_at_startup_left_to_delete_--;
113     result = 0.f;
114   }
115 
116   if (result >= kDetectThreshold) {
117     result = 1.f;
118   } else {
119     // Get proportional value.
120     // Proportion achieved with a squared raised cosine function with domain
121     // [0, kDetectThreshold) and image [0, 1), it's always increasing.
122     const float horizontal_scaling = ts::kPi / kDetectThreshold;
123     const float kHorizontalShift = ts::kPi;
124     const float kVerticalScaling = 0.5f;
125     const float kVerticalShift = 1.f;
126 
127     result = (cos(result * horizontal_scaling + kHorizontalShift)
128         + kVerticalShift) * kVerticalScaling;
129     result *= result;
130   }
131 
132   previous_results_.pop_front();
133   previous_results_.push_back(result);
134 
135   // In the current implementation we return the max of the current result and
136   // the previous results, so the high results have a width equals to
137   // |transient_length|.
138   return *std::max_element(previous_results_.begin(), previous_results_.end());
139 }
140 
141 // Looks for the highest slope and compares it with the previous ones.
142 // An exponential transformation takes this to the [0, 1] range. This value is
143 // multiplied by the detection result to avoid false positives.
ReferenceDetectionValue(const float * data,size_t length)144 float TransientDetector::ReferenceDetectionValue(const float* data,
145                                                  size_t length) {
146   if (data == NULL) {
147     using_reference_ = false;
148     return 1.f;
149   }
150   static const float kEnergyRatioThreshold = 0.2f;
151   static const float kReferenceNonLinearity = 20.f;
152   static const float kMemory = 0.99f;
153   float reference_energy = 0.f;
154   for (size_t i = 1; i < length; ++i) {
155     reference_energy += data[i] * data[i];
156   }
157   if (reference_energy == 0.f) {
158     using_reference_ = false;
159     return 1.f;
160   }
161   assert(reference_energy_ != 0);
162   float result = 1.f / (1.f + exp(kReferenceNonLinearity *
163                                   (kEnergyRatioThreshold -
164                                    reference_energy / reference_energy_)));
165   reference_energy_ =
166       kMemory * reference_energy_ + (1.f - kMemory) * reference_energy;
167 
168   using_reference_ = true;
169 
170   return result;
171 }
172 
173 }  // namespace webrtc
174