1 //
2 // Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include <catch.hpp>
7 #include <limits>
8
9 #include "DataStructures.hpp"
10 #include "Wav2LetterPreprocessor.hpp"
11
PopulateTestWavVector(std::vector<int16_t> & vec)12 void PopulateTestWavVector(std::vector<int16_t>& vec)
13 {
14 constexpr int int16max = std::numeric_limits<int16_t>::max();
15 int val = 0;
16 for (size_t i = 0; i < vec.size(); ++i, ++val)
17 {
18
19 /* We want a differential filter response from both - order 1
20 * and 2 => Don't have a linear signal here - we use a signal
21 * using squares for example. Alternate sign flips might work
22 * just as well and will be computationally less work! */
23 int valsq = val * val;
24 if (valsq > int16max)
25 {
26 val = 0;
27 valsq = 0;
28 }
29 vec[i] = valsq;
30 }
31 }
32
33 TEST_CASE("Preprocessing calculation INT8")
34 {
35 /*Test Constants: */
36 const uint32_t windowLen = 512;
37 const uint32_t windowStride = 160;
38 const float quantScale = 0.1410219967365265;
39 const int quantOffset = -11;
40 int numMfccVectors = 10;
41 const int sampFreq = 16000;
42 const int frameLenMs = 32;
43 const int frameLenSamples = sampFreq * frameLenMs * 0.001;
44 const int numMfccFeats = 13;
45 const int audioDataToPreProcess = 512 + ((numMfccVectors -1) * windowStride);
46 int outputBufferSize = numMfccVectors * numMfccFeats * 3;
47
48 /* Test wav memory */
49 std::vector <int16_t> testWav1((windowStride * numMfccVectors) +
50 (windowLen - windowStride));
51 /* Populate with dummy input */
52 PopulateTestWavVector(testWav1);
53
54 MfccParams mfccParams(sampFreq, 128, 0, 8000, numMfccFeats,
55 frameLenSamples, false, numMfccVectors);
56
57 std::unique_ptr<Wav2LetterMFCC> mfccInst = std::make_unique<Wav2LetterMFCC>(mfccParams);
58
59 std::vector<float> fullAudioData;
60
61 for(int i = 0; i < 4; ++i)
62 {
63 for (auto f : testWav1)
64 {
65 fullAudioData.emplace_back(static_cast<float>(f) / (1<<15));
66 }
67 }
68
69 Wav2LetterPreprocessor prep(frameLenSamples, windowStride, std::move(mfccInst));
70
71 std::vector<int8_t> outputBuffer(outputBufferSize);
72
73 prep.Invoke(fullAudioData.data(), audioDataToPreProcess, outputBuffer, quantOffset, quantScale);
74
75 int8_t expectedResult[numMfccVectors][numMfccFeats*3] =
76 {
77 /* Feature vec 0 */
78 -32, 4, -9, -8, -10, -10, -11, -11, -11, -11, -12, -11, -11, /* MFCCs */
79 -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, /* Delta 1 */
80 -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, /* Delta 2 */
81
82 /* Feature vec 1 */
83 -31, 4, -9, -8, -10, -10, -11, -11, -11, -11, -12, -11, -11,
84 -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
85 -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
86
87 /* Feature vec 2 */
88 -31, 4, -9, -9, -10, -10, -11, -11, -11, -11, -12, -12, -12,
89 -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
90 -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
91
92 /* Feature vec 3 */
93 -31, 4, -9, -9, -10, -10, -11, -11, -11, -11, -11, -12, -12,
94 -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
95 -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
96
97 /* Feature vec 4 : this should have valid delta 1 and delta 2 */
98 -31, 4, -9, -9, -10, -10, -11, -11, -11, -11, -11, -12, -12,
99 -38, -29, -9, 1, -2, -7, -8, -8, -12, -16, -14, -5, 5,
100 -68, -50, -13, 5, 0, -9, -9, -8, -13, -20, -19, -3, 15,
101
102 /* Feature vec 5 : this should have valid delta 1 and delta 2 */
103 -31, 4, -9, -8, -10, -10, -11, -11, -11, -11, -11, -12, -12,
104 -62, -45, -11, 5, 0, -8, -9, -8, -12, -19, -17, -3, 13,
105 -27, -22, -13, -9, -11, -12, -12, -11, -11, -13, -13, -10, -6,
106
107 /* Feature vec 6 */
108 -31, 4, -9, -8, -10, -10, -11, -11, -11, -11, -12, -11, -11,
109 -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
110 -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
111
112 /* Feature vec 7 */
113 -32, 4, -9, -8, -10, -10, -11, -11, -11, -12, -12, -11, -11,
114 -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
115 -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
116
117 /* Feature vec 8 */
118 -32, 4, -9, -8, -10, -10, -11, -11, -11, -12, -12, -11, -11,
119 -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
120 -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
121
122 /* Feature vec 9 */
123 -31, 4, -9, -8, -10, -10, -11, -11, -11, -11, -12, -11, -11,
124 -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
125 -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10
126 };
127
128 /* Check that the elements have been calculated correctly */
129 for (uint32_t j = 0; j < numMfccVectors; ++j)
130 {
131 for (uint32_t i = 0; i < numMfccFeats * 3; ++i)
132 {
133 size_t tensorIdx = (j * numMfccFeats * 3) + i;
134 CHECK(static_cast<int16_t>(outputBuffer.at(tensorIdx) == static_cast<int16_t>(expectedResult[j][i])));
135 }
136 }
137 }
138