def scan(filepath):
# sox package has an info() but it should not be used:
# it's just a wrapper that calls a set of functions
# and returns them in a dict
# ...and the functions are the wrong ones :)
info_dictionary = {
'file_type': file_info.file_type(filepath),
'sample_rate': round(file_info.sample_rate(filepath)),
'channels': file_info.channels(filepath),
'duration': file_info.duration(filepath),
# 'bit_rate': file_info.bitrate(filepath),
'encoding': file_info.encoding(filepath),
}
# bitrate is currently broken (1.37) but we can fake it
info_dictionary['bit_rate'] = round(
getsize(filepath) / info_dictionary['duration'] * 8)
# get comments too
comments = file_info.comments(filepath).splitlines()
if comments:
info_dictionary['tags'] = {}
for comment in comments:
key, value = comment.strip().split('=', 1)
info_dictionary['tags'][key.lower()] = value
return info_dictionary
def read(self, audio_metadata):
"""Read an audio file.
:param audio_metadata: metadata info of an audio
:return: raw audio data as float32 array and duration in seconds.
"""
fd = temp_path = None
# Convert it to a wav file.
if not audio_metadata.path.endswith('.wav'):
original_sample_rate = file_info.sample_rate(audio_metadata.path)
assert self._sample_rate <= original_sample_rate
transformer = Transformer()
transformer.convert(samplerate=self._sample_rate,
n_channels=self._channels,
bitdepth=self._bits_per_sample)
fd, temp_path = tempfile.mkstemp(suffix='.wav')
transformer.build(audio_metadata.path, temp_path)
if temp_path:
path = temp_path
else:
path = audio_metadata.path
# Read the audio file.
with SoundFile(path) as soundfile:
# make sure the audio properties are as expected.
assert soundfile.samplerate == self._sample_rate
assert soundfile.channels == self._channels
duration_sec = len(soundfile) / self._sample_rate
pcm = soundfile.read(dtype='float32')
# Add 0.5 second silence to the end of files containing keyword as in occasionally the user stopped
# recording right after uttering the keyword. If the detector needs some time after seeing the keyword to
# make a decision (e.g. endpointing) this is going to artificially increase the miss rates.
if audio_metadata.is_keyword:
pcm = np.append(pcm, np.zeros(self._sample_rate // 2))
if temp_path:
os.close(fd)
os.remove(temp_path)
return pcm, duration_sec
def test_empty(self):
actual = file_info.sample_rate(EMPTY_FILE)
expected = 44100
self.assertEqual(expected, actual)
def test_aiff(self):
actual = file_info.sample_rate(INPUT_FILE2)
expected = 8000
self.assertEqual(expected, actual)
def test_wav(self):
actual = file_info.sample_rate(INPUT_FILE)
expected = 44100
self.assertEqual(expected, actual)
def test_empty(self):
actual = file_info.sample_rate(EMPTY_FILE)
expected = 44100
self.assertEqual(expected, actual)
def test_aiff(self):
actual = file_info.sample_rate(INPUT_FILE2)
expected = 8000
self.assertEqual(expected, actual)
def test_wav(self):
actual = file_info.sample_rate(INPUT_FILE)
expected = 44100
self.assertEqual(expected, actual)
