def cut(input_path, output_file, metadata):
segments = metadata['segments']
segments = [segment_seconds(segment) for segment in segments]
with TempFile('.mp3') as temp_file:
# Open a new temporary file to store audio in between processes
if segments:
# Cut audio into segments and create fade in/out
# We need to use a new temporary file for each
# audio segment
temp_segments = [TempFile('.mp3') for segment in segments]
try:
for index, segment in enumerate(segments):
sox = Transformer()
sox.channels(1)
sox.norm(-24)
sox.trim(*segment)
sox.fade(1, 2, 't')
sox.build(input_path, temp_segments[index].path)
if len(segments) > 1:
# Concatenate all the audio segments back together
# and output to our main temporary file
Combiner().build(
[temp_segment.path for temp_segment in temp_segments],
temp_file.path,
'concatenate',
)
else:
# Only one segment so we don't need to combine anything
subprocess.run(
['cp', temp_segments[0].path, temp_file.path])
except Exception as e:
raise (e)
finally:
# Cleanup temporary segment files even on error
if temp_segments:
for temp_segment in temp_segments:
temp_segment.close()
# Second process: filter, compress and EQ the
# audio in temporary file and output to output_file
sox = Transformer()
sox.highpass(100)
sox.lowpass(10000)
sox.compand(0.005, 0.12, 6, [
(-90, -90),
(-70, -55),
(-50, -35),
(-32, -32),
(-24, -24),
(0, -8),
])
sox.equalizer(3000, 1000, 3)
sox.equalizer(280, 120, 3)
sox.build(temp_file.path, output_file)
def preprocess_wav(cls, fpath: Union[str, Path]) -> np.ndarray:
"""Load, resample, normalize and trim a waveform."""
transformer = Transformer()
transformer.norm()
transformer.silence(silence_threshold=1, min_silence_duration=0.1)
transformer.set_output_format(rate=cls.sample_rate,
bits=16,
channels=1)
wav = transformer.build_array(input_filepath=str(fpath))
wav = wav / (2**15)
return wav.astype(np.float32)
def loadFile(data, max_timestep):
transformer = Transformer()
transformer.norm()
# transformer.silence(silence_threshold=1, min_silence_duration=0.1)
transformer.set_output_format(rate=16000, bits=16, channels=1)
wav = transformer.build_array(input_filepath=str(data))
wav = torch.tensor(wav / (2**15)).float()
length = len(wav)
if length > max_timestep:
start = 0
end = max_timestep
length = max_timestep
wav = wav[start:end]
length = torch.tensor(length).long()
return wav, length
def loadFile_thread_exec(data):
wavs = []
lengths = []
for i in range(len(data)):
fullPath = data[i]
transformer = Transformer()
transformer.norm()
transformer.silence(silence_threshold=1, min_silence_duration=0.1)
transformer.set_output_format(rate=16000, bits=16, channels=1)
wav = transformer.build_array(input_filepath=str(fullPath))
wav = torch.tensor(wav / (2**15)).float()
length = len(wav)
if length > max_timestep:
start = random.randint(0, int(length - max_timestep))
end = start + max_timestep
length = max_timestep
wav = wav[start:end]
wavs.append(wav)
lengths.append(torch.tensor(length).long())
return wavs, lengths
def _processSamples(sample_list):
for sample in sample_list:
sample_new_name = _renameSample(sample)
_out = join(out_path, sample_new_name)
processed_samples.append(_out)
_in = sample
# Sox processing using Transform instance
tfm = Transformer()
tfm.convert(samplerate=44100, n_channels=2, bitdepth=16)
if NORMALIZE:
tfm.norm(db_level=-3)
if SILENCE:
tfm.silence(location=-1,
silence_threshold=0.05,
min_silence_duration=0.1)
if PADDING:
tfm.pad(0, PADDING)
tfm.build(_in, _out)
