xref: /aosp_15_r20/external/armnn/python/pyarmnn/examples/speech_recognition/run_audio_file.py (revision 89c4ff92f2867872bb9e2354d150bf0c8c502810)
1# Copyright © 2021 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
6import sys
7import os
8import numpy as np
9
10script_dir = os.path.dirname(__file__)
11sys.path.insert(1, os.path.join(script_dir, '..', 'common'))
12
13from argparse import ArgumentParser
14from network_executor import ArmnnNetworkExecutor
15from utils import prepare_input_data
16from audio_capture import AudioCaptureParams, capture_audio
17from audio_utils import decode_text, display_text
18from wav2letter_mfcc import Wav2LetterMFCC, W2LAudioPreprocessor
19from mfcc import MFCCParams
20
21# Model Specific Labels
22labels = {0: 'a', 1: 'b', 2: 'c', 3: 'd', 4: 'e', 5: 'f', 6: 'g', 7: 'h', 8: 'i', 9: 'j', 10: 'k', 11: 'l', 12: 'm',
23          13: 'n',
24          14: 'o', 15: 'p', 16: 'q', 17: 'r', 18: 's', 19: 't', 20: 'u', 21: 'v', 22: 'w', 23: 'x', 24: 'y',
25          25: 'z',
26          26: "'", 27: ' ', 28: '$'}
27
28
29def parse_args():
30    parser = ArgumentParser(description="ASR with PyArmNN")
31    parser.add_argument(
32        "--audio_file_path",
33        required=True,
34        type=str,
35        help="Path to the audio file to perform ASR",
36    )
37    parser.add_argument(
38        "--model_file_path",
39        required=True,
40        type=str,
41        help="Path to ASR model to use",
42    )
43    parser.add_argument(
44        "--preferred_backends",
45        type=str,
46        nargs="+",
47        default=["CpuAcc", "CpuRef"],
48        help="""List of backends in order of preference for optimizing
49        subgraphs, falling back to the next backend in the list on unsupported
50        layers. Defaults to [CpuAcc, CpuRef]""",
51    )
52    return parser.parse_args()
53
54
55def main(args):
56    # Read command line args
57    audio_file = args.audio_file_path
58
59    # Create the ArmNN inference runner
60    network = ArmnnNetworkExecutor(args.model_file_path, args.preferred_backends)
61
62    # Specify model specific audio data requirements
63    audio_capture_params = AudioCaptureParams(dtype=np.float32, overlap=31712, min_samples=47712, sampling_freq=16000,
64                                              mono=True)
65
66    buffer = capture_audio(audio_file, audio_capture_params)
67
68    # Extract features and create the preprocessor
69
70    mfcc_params = MFCCParams(sampling_freq=16000, num_fbank_bins=128, mel_lo_freq=0, mel_hi_freq=8000,
71                             num_mfcc_feats=13, frame_len=512, use_htk_method=False, n_fft=512)
72
73    wmfcc = Wav2LetterMFCC(mfcc_params)
74    preprocessor = W2LAudioPreprocessor(wmfcc, model_input_size=296, stride=160)
75    current_r_context = ""
76    is_first_window = True
77
78    print("Processing Audio Frames...")
79    for audio_data in buffer:
80        # Prepare the input Tensors
81        input_data = prepare_input_data(audio_data, network.get_data_type(), network.get_input_quantization_scale(0),
82                                        network.get_input_quantization_offset(0), preprocessor)
83
84        # Run inference
85        output_result = network.run([input_data])
86
87        # Slice and Decode the text, and store the right context
88        current_r_context, text = decode_text(is_first_window, labels, output_result)
89
90        is_first_window = False
91
92        display_text(text)
93
94    print(current_r_context, flush=True)
95
96
97if __name__ == "__main__":
98    args = parse_args()
99    main(args)
100