|
1
|
|
|
import os |
|
2
|
|
|
|
|
3
|
|
|
import h5py |
|
4
|
|
|
import scipy.io.wavfile |
|
5
|
|
|
|
|
6
|
|
|
from fuel.converters.base import fill_hdf5_file |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
def convert_youtube_audio(directory, output_directory, youtube_id, channels, |
|
10
|
|
|
sample, output_filename=None): |
|
11
|
|
|
input_file = os.path.join(directory, '{}.m4a'.format(youtube_id)) |
|
12
|
|
|
wav_filename = '{}.wav'.format(youtube_id) |
|
13
|
|
|
wav_file = os.path.join(directory, wav_filename) |
|
14
|
|
|
command = "ffmpeg -y -i {} -ac {} -ar {} {}".format( |
|
15
|
|
|
input_file, channels, sample, wav_file) |
|
16
|
|
|
os.system(command) |
|
17
|
|
|
|
|
18
|
|
|
# Load WAV into array |
|
19
|
|
|
_, data = scipy.io.wavfile.read(wav_file) |
|
20
|
|
|
if data.ndim == 1: |
|
21
|
|
|
data = data[:, None] |
|
22
|
|
|
|
|
23
|
|
|
# Store in HDF5 |
|
24
|
|
|
if output_filename is None: |
|
25
|
|
|
output_filename = '{}.hdf5'.format(youtube_id) |
|
26
|
|
|
output_file = os.path.join(output_directory, output_filename) |
|
27
|
|
|
|
|
28
|
|
|
with h5py.File(output_file, 'w') as h5file: |
|
29
|
|
|
fill_hdf5_file(h5file, (('train', 'features', data),)) |
|
30
|
|
|
h5file['features'].dims[0].label = 'time' |
|
31
|
|
|
h5file['features'].dims[1].label = 'feature' |
|
32
|
|
|
|
|
33
|
|
|
return (output_file,) |
|
34
|
|
|
|
|
35
|
|
|
|
|
36
|
|
|
def fill_subparser(subparser): |
|
37
|
|
|
subparser.add_argument( |
|
38
|
|
|
'--youtube-id', type=str, required=True, |
|
39
|
|
|
help=("The YouTube ID of the video from which to extract audio, " |
|
40
|
|
|
"usually an 11-character string.") |
|
41
|
|
|
) |
|
42
|
|
|
subparser.add_argument( |
|
43
|
|
|
'--channels', type=int, default=1, |
|
44
|
|
|
help="The number of audio channels to convert to" |
|
45
|
|
|
) |
|
46
|
|
|
subparser.add_argument( |
|
47
|
|
|
'--sample', type=int, default=16000, |
|
48
|
|
|
help="The sampling rate in Hz" |
|
49
|
|
|
) |
|
50
|
|
|
return convert_youtube_audio |
|
51
|
|
|
|
mila-udem /
fuel
