1 /* 2 * Copyright (c) 2014 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 <cmath> 12 #include <algorithm> 13 #include <vector> 14 15 #include "testing/gtest/include/gtest/gtest.h" 16 #include "webrtc/base/arraysize.h" 17 #include "webrtc/base/format_macros.h" 18 #include "webrtc/base/scoped_ptr.h" 19 #include "webrtc/common_audio/audio_converter.h" 20 #include "webrtc/common_audio/channel_buffer.h" 21 #include "webrtc/common_audio/resampler/push_sinc_resampler.h" 22 23 namespace webrtc { 24 25 typedef rtc::scoped_ptr<ChannelBuffer<float>> ScopedBuffer; 26 27 // Sets the signal value to increase by |data| with every sample. 28 ScopedBuffer CreateBuffer(const std::vector<float>& data, size_t frames) { 29 const size_t num_channels = data.size(); 30 ScopedBuffer sb(new ChannelBuffer<float>(frames, num_channels)); 31 for (size_t i = 0; i < num_channels; ++i) 32 for (size_t j = 0; j < frames; ++j) 33 sb->channels()[i][j] = data[i] * j; 34 return sb; 35 } 36 37 void VerifyParams(const ChannelBuffer<float>& ref, 38 const ChannelBuffer<float>& test) { 39 EXPECT_EQ(ref.num_channels(), test.num_channels()); 40 EXPECT_EQ(ref.num_frames(), test.num_frames()); 41 } 42 43 // Computes the best SNR based on the error between |ref_frame| and 44 // |test_frame|. It searches around |expected_delay| in samples between the 45 // signals to compensate for the resampling delay. 46 float ComputeSNR(const ChannelBuffer<float>& ref, 47 const ChannelBuffer<float>& test, 48 size_t expected_delay) { 49 VerifyParams(ref, test); 50 float best_snr = 0; 51 size_t best_delay = 0; 52 53 // Search within one sample of the expected delay. 54 for (size_t delay = std::max(expected_delay, static_cast<size_t>(1)) - 1; 55 delay <= std::min(expected_delay + 1, ref.num_frames()); 56 ++delay) { 57 float mse = 0; 58 float variance = 0; 59 float mean = 0; 60 for (size_t i = 0; i < ref.num_channels(); ++i) { 61 for (size_t j = 0; j < ref.num_frames() - delay; ++j) { 62 float error = ref.channels()[i][j] - test.channels()[i][j + delay]; 63 mse += error * error; 64 variance += ref.channels()[i][j] * ref.channels()[i][j]; 65 mean += ref.channels()[i][j]; 66 } 67 } 68 69 const size_t length = ref.num_channels() * (ref.num_frames() - delay); 70 mse /= length; 71 variance /= length; 72 mean /= length; 73 variance -= mean * mean; 74 float snr = 100; // We assign 100 dB to the zero-error case. 75 if (mse > 0) 76 snr = 10 * std::log10(variance / mse); 77 if (snr > best_snr) { 78 best_snr = snr; 79 best_delay = delay; 80 } 81 } 82 printf("SNR=%.1f dB at delay=%" PRIuS "\n", best_snr, best_delay); 83 return best_snr; 84 } 85 86 // Sets the source to a linearly increasing signal for which we can easily 87 // generate a reference. Runs the AudioConverter and ensures the output has 88 // sufficiently high SNR relative to the reference. 89 void RunAudioConverterTest(size_t src_channels, 90 int src_sample_rate_hz, 91 size_t dst_channels, 92 int dst_sample_rate_hz) { 93 const float kSrcLeft = 0.0002f; 94 const float kSrcRight = 0.0001f; 95 const float resampling_factor = (1.f * src_sample_rate_hz) / 96 dst_sample_rate_hz; 97 const float dst_left = resampling_factor * kSrcLeft; 98 const float dst_right = resampling_factor * kSrcRight; 99 const float dst_mono = (dst_left + dst_right) / 2; 100 const size_t src_frames = static_cast<size_t>(src_sample_rate_hz / 100); 101 const size_t dst_frames = static_cast<size_t>(dst_sample_rate_hz / 100); 102 103 std::vector<float> src_data(1, kSrcLeft); 104 if (src_channels == 2) 105 src_data.push_back(kSrcRight); 106 ScopedBuffer src_buffer = CreateBuffer(src_data, src_frames); 107 108 std::vector<float> dst_data(1, 0); 109 std::vector<float> ref_data; 110 if (dst_channels == 1) { 111 if (src_channels == 1) 112 ref_data.push_back(dst_left); 113 else 114 ref_data.push_back(dst_mono); 115 } else { 116 dst_data.push_back(0); 117 ref_data.push_back(dst_left); 118 if (src_channels == 1) 119 ref_data.push_back(dst_left); 120 else 121 ref_data.push_back(dst_right); 122 } 123 ScopedBuffer dst_buffer = CreateBuffer(dst_data, dst_frames); 124 ScopedBuffer ref_buffer = CreateBuffer(ref_data, dst_frames); 125 126 // The sinc resampler has a known delay, which we compute here. 127 const size_t delay_frames = src_sample_rate_hz == dst_sample_rate_hz ? 0 : 128 static_cast<size_t>( 129 PushSincResampler::AlgorithmicDelaySeconds(src_sample_rate_hz) * 130 dst_sample_rate_hz); 131 // SNR reported on the same line later. 132 printf("(%" PRIuS ", %d Hz) -> (%" PRIuS ", %d Hz) ", 133 src_channels, src_sample_rate_hz, dst_channels, dst_sample_rate_hz); 134 135 rtc::scoped_ptr<AudioConverter> converter = AudioConverter::Create( 136 src_channels, src_frames, dst_channels, dst_frames); 137 converter->Convert(src_buffer->channels(), src_buffer->size(), 138 dst_buffer->channels(), dst_buffer->size()); 139 140 EXPECT_LT(43.f, 141 ComputeSNR(*ref_buffer.get(), *dst_buffer.get(), delay_frames)); 142 } 143 144 TEST(AudioConverterTest, ConversionsPassSNRThreshold) { 145 const int kSampleRates[] = {8000, 16000, 32000, 44100, 48000}; 146 const size_t kChannels[] = {1, 2}; 147 for (size_t src_rate = 0; src_rate < arraysize(kSampleRates); ++src_rate) { 148 for (size_t dst_rate = 0; dst_rate < arraysize(kSampleRates); ++dst_rate) { 149 for (size_t src_channel = 0; src_channel < arraysize(kChannels); 150 ++src_channel) { 151 for (size_t dst_channel = 0; dst_channel < arraysize(kChannels); 152 ++dst_channel) { 153 RunAudioConverterTest(kChannels[src_channel], kSampleRates[src_rate], 154 kChannels[dst_channel], kSampleRates[dst_rate]); 155 } 156 } 157 } 158 } 159 } 160 161 } // namespace webrtc 162