|
| 1 | +// sherpa-onnx/csrc/sherpa-onnx-vad-with-online-asr.cc |
| 2 | +// |
| 3 | +// Copyright (c) 2025 Xiaomi Corporation |
| 4 | +// Copyright (c) 2025 Pingfeng Luo |
| 5 | +// |
| 6 | +// This file demonstrates how to use vad in streaming speech recognition |
| 7 | +// |
| 8 | + |
| 9 | +#include <stdio.h> |
| 10 | + |
| 11 | +#include <chrono> // NOLINT |
| 12 | +#include <string> |
| 13 | +#include <vector> |
| 14 | + |
| 15 | +#include "sherpa-onnx/csrc/online-recognizer.h" |
| 16 | +#include "sherpa-onnx/csrc/online-stream.h" |
| 17 | +#include "sherpa-onnx/csrc/parse-options.h" |
| 18 | +#include "sherpa-onnx/csrc/resample.h" |
| 19 | +#include "sherpa-onnx/csrc/symbol-table.h" |
| 20 | +#include "sherpa-onnx/csrc/voice-activity-detector.h" |
| 21 | +#include "sherpa-onnx/csrc/wave-reader.h" |
| 22 | + |
| 23 | +int32_t main(int32_t argc, char *argv[]) { |
| 24 | + const char *kUsageMessage = R"usage( |
| 25 | +Speech recognition using VAD + streaming models with sherpa-onnx-vad-with-online-asr. |
| 26 | +This is useful when testing long audio. |
| 27 | +
|
| 28 | +Usage: |
| 29 | +
|
| 30 | +Note you can download silero_vad.onnx using |
| 31 | +
|
| 32 | +wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx |
| 33 | +
|
| 34 | +(1) Streaming transducer |
| 35 | +
|
| 36 | + ./bin/sherpa-onnx-vad-with-online-asr \ |
| 37 | + --silero-vad-model=/path/to/silero_vad.onnx \ |
| 38 | + --tokens=/path/to/tokens.txt \ |
| 39 | + --encoder=/path/to/encoder.onnx \ |
| 40 | + --decoder=/path/to/decoder.onnx \ |
| 41 | + --joiner=/path/to/joiner.onnx \ |
| 42 | + --provider=cpu \ |
| 43 | + --num-threads=2 \ |
| 44 | + --decoding-method=greedy_search \ |
| 45 | + /path/to/long_duration.wav |
| 46 | +
|
| 47 | +(2) Streaming zipformer2 CTC |
| 48 | +
|
| 49 | + wget -q https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-streaming-zipformer-ctc-multi-zh-hans-2023-12-13.tar.bz2 |
| 50 | + tar xvf sherpa-onnx-streaming-zipformer-ctc-multi-zh-hans-2023-12-13.tar.bz2 |
| 51 | +
|
| 52 | + ./bin/sherpa-onnx-vad-with-online-asr \ |
| 53 | + --debug=1 \ |
| 54 | + --silero-vad-model=/path/to/silero_vad.onnx \ |
| 55 | + --zipformer2-ctc-model=./sherpa-onnx-streaming-zipformer-ctc-multi-zh-hans-2023-12-13/ctc-epoch-20-avg-1-chunk-16-left-128.onnx \ |
| 56 | + --tokens=./sherpa-onnx-streaming-zipformer-ctc-multi-zh-hans-2023-12-13/tokens.txt \ |
| 57 | + ./sherpa-onnx-streaming-zipformer-ctc-multi-zh-hans-2023-12-13/test_wavs/DEV_T0000000000.wav |
| 58 | +
|
| 59 | +(3) Streaming paraformer |
| 60 | +
|
| 61 | + wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-streaming-paraformer-bilingual-zh-en.tar.bz2 |
| 62 | + tar xvf sherpa-onnx-streaming-paraformer-bilingual-zh-en.tar.bz2 |
| 63 | +
|
| 64 | + ./bin/sherpa-onnx-vad-with-online-asr \ |
| 65 | + --silero-vad-model=/path/to/silero_vad.onnx \ |
| 66 | + --tokens=./sherpa-onnx-streaming-paraformer-bilingual-zh-en/tokens.txt \ |
| 67 | + --paraformer-encoder=./sherpa-onnx-streaming-paraformer-bilingual-zh-en/encoder.onnx \ |
| 68 | + --paraformer-decoder=./sherpa-onnx-streaming-paraformer-bilingual-zh-en/decoder.onnx \ |
| 69 | + /path/to/long_duration.wav |
| 70 | +
|
| 71 | +
|
| 72 | +The input wav should be of single channel, 16-bit PCM encoded wave file; its |
| 73 | +sampling rate can be arbitrary and does not need to be 16kHz. |
| 74 | +
|
| 75 | +Please refer to |
| 76 | +https://k2-fsa.github.io/sherpa/onnx/pretrained_models/index.html |
| 77 | +for a list of pre-trained models to download. |
| 78 | +)usage"; |
| 79 | + |
| 80 | + sherpa_onnx::ParseOptions po(kUsageMessage); |
| 81 | + sherpa_onnx::OnlineRecognizerConfig asr_config; |
| 82 | + asr_config.Register(&po); |
| 83 | + |
| 84 | + sherpa_onnx::VadModelConfig vad_config; |
| 85 | + vad_config.Register(&po); |
| 86 | + |
| 87 | + po.Read(argc, argv); |
| 88 | + if (po.NumArgs() != 1) { |
| 89 | + fprintf(stderr, "Error: Please provide exactly 1 wave file. Given: %d\n\n", |
| 90 | + po.NumArgs()); |
| 91 | + po.PrintUsage(); |
| 92 | + exit(EXIT_FAILURE); |
| 93 | + } |
| 94 | + |
| 95 | + fprintf(stderr, "%s\n", vad_config.ToString().c_str()); |
| 96 | + fprintf(stderr, "%s\n", asr_config.ToString().c_str()); |
| 97 | + |
| 98 | + if (!vad_config.Validate()) { |
| 99 | + fprintf(stderr, "Errors in vad_config!\n"); |
| 100 | + return -1; |
| 101 | + } |
| 102 | + |
| 103 | + if (!asr_config.Validate()) { |
| 104 | + fprintf(stderr, "Errors in ASR config!\n"); |
| 105 | + return -1; |
| 106 | + } |
| 107 | + |
| 108 | + fprintf(stderr, "Creating recognizer ...\n"); |
| 109 | + sherpa_onnx::OnlineRecognizer recognizer(asr_config); |
| 110 | + fprintf(stderr, "Recognizer created!\n"); |
| 111 | + |
| 112 | + auto vad = std::make_unique<sherpa_onnx::VoiceActivityDetector>(vad_config); |
| 113 | + |
| 114 | + fprintf(stderr, "Started\n"); |
| 115 | + const auto begin = std::chrono::steady_clock::now(); |
| 116 | + |
| 117 | + std::string wave_filename = po.GetArg(1); |
| 118 | + fprintf(stderr, "Reading: %s\n", wave_filename.c_str()); |
| 119 | + int32_t sampling_rate = -1; |
| 120 | + bool is_ok = false; |
| 121 | + auto samples = sherpa_onnx::ReadWave(wave_filename, &sampling_rate, &is_ok); |
| 122 | + if (!is_ok) { |
| 123 | + fprintf(stderr, "Failed to read '%s'\n", wave_filename.c_str()); |
| 124 | + return -1; |
| 125 | + } |
| 126 | + |
| 127 | + if (sampling_rate != 16000) { |
| 128 | + fprintf(stderr, "Resampling from %d Hz to 16000 Hz\n", sampling_rate); |
| 129 | + float min_freq = std::min(sampling_rate, 16000) |
| 130 | + float lowpass_cutoff = 0.99 * 0.5 * min_freq; |
| 131 | + |
| 132 | + int32_t lowpass_filter_width = 6; |
| 133 | + auto resampler = std::make_unique<sherpa_onnx::LinearResample>( |
| 134 | + sampling_rate, 16000, lowpass_cutoff, lowpass_filter_width); |
| 135 | + std::vector<float> out_samples; |
| 136 | + resampler->Resample(samples.data(), samples.size(), true, &out_samples); |
| 137 | + samples = std::move(out_samples); |
| 138 | + fprintf(stderr, "Resampling done\n"); |
| 139 | + } |
| 140 | + |
| 141 | + fprintf(stderr, "Started!\n"); |
| 142 | + int32_t window_size = vad_config.ten_vad.model.empty() |
| 143 | + ? vad_config.silero_vad.window_size : vad_config.ten_vad.window_size; |
| 144 | + int32_t offset = 0; |
| 145 | + int32_t segment_id = 0; |
| 146 | + bool speech_started = false; |
| 147 | + while (offset < samples.size()) { |
| 148 | + if (offset + window_size <= samples.size()) { |
| 149 | + vad->AcceptWaveform(samples.data() + offset, window_size); |
| 150 | + } else { |
| 151 | + vad->Flush(); |
| 152 | + } |
| 153 | + offset += window_size; |
| 154 | + if (vad->IsSpeechDetected() && !speech_started) { |
| 155 | + // new voice activity |
| 156 | + speech_started = true; |
| 157 | + segment_id++; |
| 158 | + } else if (!vad->IsSpeechDetected() && speech_started) { |
| 159 | + // end voice activity |
| 160 | + speech_started = false; |
| 161 | + } |
| 162 | + |
| 163 | + while (!vad->Empty()) { |
| 164 | + const auto &segment = vad->Front(); |
| 165 | + float duration = segment.samples.size() / 16000.; |
| 166 | + float start_time = segment.start / 16000.; |
| 167 | + float end_time = start_time + duration; |
| 168 | + auto s = recognizer.CreateStream(); |
| 169 | + s->AcceptWaveform(16000, segment.samples.data(), segment.samples.size()); |
| 170 | + s->InputFinished(); |
| 171 | + while (recognizer.IsReady(s.get())) { |
| 172 | + recognizer.DecodeStream(s.get()); |
| 173 | + } |
| 174 | + auto text = recognizer.GetResult(s.get()).text; |
| 175 | + if (!text.empty()) { |
| 176 | + fprintf(stderr, "vad segment(%d:%.3f-%.3f) results: %s\n", |
| 177 | + segment_id, start_time, end_time, text.c_str()); |
| 178 | + } |
| 179 | + vad->Pop(); |
| 180 | + } |
| 181 | + } |
| 182 | + |
| 183 | + const auto end = std::chrono::steady_clock::now(); |
| 184 | + |
| 185 | + float elapsed_seconds = |
| 186 | + std::chrono::duration_cast<std::chrono::milliseconds>(end - begin) |
| 187 | + .count() / |
| 188 | + 1000.; |
| 189 | + |
| 190 | + fprintf(stderr, "num threads: %d\n", asr_config.model_config.num_threads); |
| 191 | + fprintf(stderr, "decoding method: %s\n", asr_config.decoding_method.c_str()); |
| 192 | + if (asr_config.decoding_method == "modified_beam_search") { |
| 193 | + fprintf(stderr, "max active paths: %d\n", asr_config.max_active_paths); |
| 194 | + } |
| 195 | + |
| 196 | + float duration = samples.size() / 16000.; |
| 197 | + fprintf(stderr, "Elapsed seconds: %.3f s\n", elapsed_seconds); |
| 198 | + float rtf = elapsed_seconds / duration; |
| 199 | + fprintf(stderr, "Real time factor (RTF): %.3f / %.3f = %.3f\n", |
| 200 | + elapsed_seconds, duration, rtf); |
| 201 | + |
| 202 | + return 0; |
| 203 | +} |
0 commit comments