Éanna Ó Catháin | c6ab02a | 2021-04-07 14:35:25 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | #include <iostream> |
| 6 | #include <map> |
| 7 | #include <vector> |
| 8 | #include <algorithm> |
| 9 | #include <cmath> |
| 10 | |
| 11 | #include "CmdArgsParser.hpp" |
| 12 | #include "ArmnnNetworkExecutor.hpp" |
| 13 | #include "AudioCapture.hpp" |
| 14 | #include "Preprocess.hpp" |
| 15 | #include "Decoder.hpp" |
| 16 | #include "SpeechRecognitionPipeline.hpp" |
| 17 | |
| 18 | |
| 19 | using InferenceResult = std::vector<int8_t>; |
| 20 | using InferenceResults = std::vector<InferenceResult>; |
| 21 | |
| 22 | const std::string AUDIO_FILE_PATH = "--audio-file-path"; |
| 23 | const std::string MODEL_FILE_PATH = "--model-file-path"; |
| 24 | const std::string LABEL_PATH = "--label-path"; |
| 25 | const std::string PREFERRED_BACKENDS = "--preferred-backends"; |
| 26 | const std::string HELP = "--help"; |
| 27 | |
| 28 | std::map<int, std::string> labels = { |
| 29 | {0, "a" }, |
| 30 | {1, "b" }, |
| 31 | {2, "c" }, |
| 32 | {3, "d" }, |
| 33 | {4, "e" }, |
| 34 | {5, "f" }, |
| 35 | {6, "g" }, |
| 36 | {7, "h" }, |
| 37 | {8, "i" }, |
| 38 | {9, "j" }, |
| 39 | {10,"k" }, |
| 40 | {11,"l" }, |
| 41 | {12,"m" }, |
| 42 | {13,"n" }, |
| 43 | {14,"o" }, |
| 44 | {15,"p" }, |
| 45 | {16,"q" }, |
| 46 | {17,"r" }, |
| 47 | {18,"s" }, |
| 48 | {19,"t" }, |
| 49 | {20,"u" }, |
| 50 | {21,"v" }, |
| 51 | {22,"w" }, |
| 52 | {23,"x" }, |
| 53 | {24,"y" }, |
| 54 | {25,"z" }, |
| 55 | {26, "\'" }, |
| 56 | {27, " "}, |
| 57 | {28,"$" } |
| 58 | }; |
| 59 | |
| 60 | /* |
| 61 | * The accepted options for this Speech Recognition executable |
| 62 | */ |
| 63 | static std::map<std::string, std::string> CMD_OPTIONS = { |
| 64 | {AUDIO_FILE_PATH, "[REQUIRED] Path to the Audio file to run speech recognition on"}, |
| 65 | {MODEL_FILE_PATH, "[REQUIRED] Path to the Speech Recognition model to use"}, |
| 66 | {PREFERRED_BACKENDS, "[OPTIONAL] Takes the preferred backends in preference order, separated by comma." |
| 67 | " For example: CpuAcc,GpuAcc,CpuRef. Accepted options: [CpuAcc, CpuRef, GpuAcc]." |
| 68 | " Defaults to CpuAcc,CpuRef"} |
| 69 | }; |
| 70 | |
| 71 | /* |
| 72 | * Reads the user supplied backend preference, splits it by comma, and returns an ordered vector |
| 73 | */ |
| 74 | std::vector<armnn::BackendId> GetPreferredBackendList(const std::string& preferredBackends) |
| 75 | { |
| 76 | std::vector<armnn::BackendId> backends; |
| 77 | std::stringstream ss(preferredBackends); |
| 78 | |
| 79 | while(ss.good()) |
| 80 | { |
| 81 | std::string backend; |
| 82 | std::getline( ss, backend, ',' ); |
| 83 | backends.emplace_back(backend); |
| 84 | } |
| 85 | return backends; |
| 86 | } |
| 87 | |
| 88 | int main(int argc, char *argv[]) |
| 89 | { |
| 90 | // Wav2Letter ASR SETTINGS |
| 91 | int SAMP_FREQ = 16000; |
| 92 | int FRAME_LEN_MS = 32; |
| 93 | int FRAME_LEN_SAMPLES = SAMP_FREQ * FRAME_LEN_MS * 0.001; |
| 94 | int NUM_MFCC_FEATS = 13; |
| 95 | int MFCC_WINDOW_LEN = 512; |
| 96 | int MFCC_WINDOW_STRIDE = 160; |
| 97 | const int NUM_MFCC_VECTORS = 296; |
| 98 | int SAMPLES_PER_INFERENCE = MFCC_WINDOW_LEN + ((NUM_MFCC_VECTORS -1) * MFCC_WINDOW_STRIDE); |
| 99 | int MEL_LO_FREQ = 0; |
| 100 | int MEL_HI_FREQ = 8000; |
| 101 | int NUM_FBANK_BIN = 128; |
| 102 | int INPUT_WINDOW_LEFT_CONTEXT = 98; |
| 103 | int INPUT_WINDOW_RIGHT_CONTEXT = 98; |
| 104 | int INPUT_WINDOW_INNER_CONTEXT = NUM_MFCC_VECTORS - |
| 105 | (INPUT_WINDOW_LEFT_CONTEXT + INPUT_WINDOW_RIGHT_CONTEXT); |
| 106 | int SLIDING_WINDOW_OFFSET = INPUT_WINDOW_INNER_CONTEXT * MFCC_WINDOW_STRIDE; |
| 107 | |
| 108 | |
| 109 | MfccParams mfccParams(SAMP_FREQ, NUM_FBANK_BIN, |
| 110 | MEL_LO_FREQ, MEL_HI_FREQ, NUM_MFCC_FEATS, FRAME_LEN_SAMPLES, false, NUM_MFCC_VECTORS); |
| 111 | |
| 112 | MFCC mfccInst = MFCC(mfccParams); |
| 113 | |
| 114 | Preprocess preprocessor(MFCC_WINDOW_LEN, MFCC_WINDOW_STRIDE, mfccInst); |
| 115 | |
| 116 | bool isFirstWindow = true; |
| 117 | std::string currentRContext = ""; |
| 118 | |
| 119 | std::map <std::string, std::string> options; |
| 120 | |
| 121 | int result = ParseOptions(options, CMD_OPTIONS, argv, argc); |
| 122 | if (result != 0) |
| 123 | { |
| 124 | return result; |
| 125 | } |
| 126 | |
| 127 | // Create the network options |
| 128 | common::PipelineOptions pipelineOptions; |
| 129 | pipelineOptions.m_ModelFilePath = GetSpecifiedOption(options, MODEL_FILE_PATH); |
| 130 | |
| 131 | if (CheckOptionSpecified(options, PREFERRED_BACKENDS)) |
| 132 | { |
| 133 | pipelineOptions.m_backends = GetPreferredBackendList((GetSpecifiedOption(options, PREFERRED_BACKENDS))); |
| 134 | } |
| 135 | else |
| 136 | { |
| 137 | pipelineOptions.m_backends = {"CpuAcc", "CpuRef"}; |
| 138 | } |
| 139 | |
| 140 | asr::IPipelinePtr asrPipeline = asr::CreatePipeline(pipelineOptions, labels); |
| 141 | |
| 142 | asr::AudioCapture capture; |
| 143 | std::vector<float> audioData = capture.LoadAudioFile(GetSpecifiedOption(options, AUDIO_FILE_PATH)); |
| 144 | capture.InitSlidingWindow(audioData.data(), audioData.size(), SAMPLES_PER_INFERENCE, SLIDING_WINDOW_OFFSET); |
| 145 | |
| 146 | while (capture.HasNext()) |
| 147 | { |
| 148 | std::vector<float> audioBlock = capture.Next(); |
| 149 | InferenceResults results; |
| 150 | |
| 151 | std::vector<int8_t> preprocessedData = asrPipeline->PreProcessing<float, int8_t>(audioBlock, preprocessor); |
| 152 | asrPipeline->Inference<int8_t>(preprocessedData, results); |
| 153 | asrPipeline->PostProcessing<int8_t>(results, isFirstWindow, !capture.HasNext(), currentRContext); |
| 154 | } |
| 155 | |
| 156 | return 0; |
| 157 | } |