Éanna Ó Catháin | 145c88f | 2020-11-16 14:12:11 +0000 | [diff] [blame] | 1 | # Copyright © 2020 Arm Ltd and Contributors. All rights reserved. |
| 2 | # SPDX-License-Identifier: MIT |
| 3 | |
| 4 | """Automatic speech recognition with PyArmNN demo for processing audio clips to text.""" |
| 5 | |
| 6 | import sys |
| 7 | import os |
| 8 | from argparse import ArgumentParser |
| 9 | |
| 10 | script_dir = os.path.dirname(__file__) |
| 11 | sys.path.insert(1, os.path.join(script_dir, '..', 'common')) |
| 12 | |
| 13 | from network_executor import ArmnnNetworkExecutor |
| 14 | from utils import dict_labels |
| 15 | from preprocess import MFCCParams, Preprocessor, MFCC |
| 16 | from audio_capture import AudioCapture, ModelParams |
| 17 | from audio_utils import decode_text, prepare_input_tensors, display_text |
| 18 | |
| 19 | |
| 20 | def parse_args(): |
| 21 | parser = ArgumentParser(description="ASR with PyArmNN") |
| 22 | parser.add_argument( |
| 23 | "--audio_file_path", |
| 24 | required=True, |
| 25 | type=str, |
| 26 | help="Path to the audio file to perform ASR", |
| 27 | ) |
| 28 | parser.add_argument( |
| 29 | "--model_file_path", |
| 30 | required=True, |
| 31 | type=str, |
| 32 | help="Path to ASR model to use", |
| 33 | ) |
| 34 | parser.add_argument( |
| 35 | "--labels_file_path", |
| 36 | required=True, |
| 37 | type=str, |
| 38 | help="Path to text file containing labels to map to model output", |
| 39 | ) |
| 40 | parser.add_argument( |
| 41 | "--preferred_backends", |
| 42 | type=str, |
| 43 | nargs="+", |
| 44 | default=["CpuAcc", "CpuRef"], |
| 45 | help="""List of backends in order of preference for optimizing |
| 46 | subgraphs, falling back to the next backend in the list on unsupported |
| 47 | layers. Defaults to [CpuAcc, CpuRef]""", |
| 48 | ) |
| 49 | return parser.parse_args() |
| 50 | |
| 51 | |
| 52 | def main(args): |
| 53 | # Read command line args |
| 54 | audio_file = args.audio_file_path |
| 55 | model = ModelParams(args.model_file_path) |
| 56 | labels = dict_labels(args.labels_file_path) |
| 57 | |
| 58 | # Create the ArmNN inference runner |
| 59 | network = ArmnnNetworkExecutor(model.path, args.preferred_backends) |
| 60 | |
| 61 | audio_capture = AudioCapture(model) |
| 62 | buffer = audio_capture.from_audio_file(audio_file) |
| 63 | |
| 64 | # Create the preprocessor |
| 65 | mfcc_params = MFCCParams(sampling_freq=16000, num_fbank_bins=128, mel_lo_freq=0, mel_hi_freq=8000, |
| 66 | num_mfcc_feats=13, frame_len=512, use_htk_method=False, n_FFT=512) |
| 67 | mfcc = MFCC(mfcc_params) |
| 68 | preprocessor = Preprocessor(mfcc, model_input_size=1044, stride=160) |
| 69 | |
| 70 | text = "" |
| 71 | current_r_context = "" |
| 72 | is_first_window = True |
| 73 | |
| 74 | print("Processing Audio Frames...") |
| 75 | for audio_data in buffer: |
| 76 | # Prepare the input Tensors |
| 77 | input_tensors = prepare_input_tensors(audio_data, network.input_binding_info, preprocessor) |
| 78 | |
| 79 | # Run inference |
| 80 | output_result = network.run(input_tensors) |
| 81 | |
| 82 | # Slice and Decode the text, and store the right context |
| 83 | current_r_context, text = decode_text(is_first_window, labels, output_result) |
| 84 | |
| 85 | is_first_window = False |
| 86 | |
| 87 | display_text(text) |
| 88 | |
| 89 | print(current_r_context, flush=True) |
| 90 | |
| 91 | |
| 92 | if __name__ == "__main__": |
| 93 | args = parse_args() |
| 94 | main(args) |