def preprocess(filename, timidity, latency, truncate, pad=1, get_raw=False):
"""
Preprocess an audio file ands its MIDI counterpart. Computes transforms and labels.
:param filename: audio filename
:param timidity: set to True if the files was rendered with timidity
:param latency: in seconds
:param truncate: in seconds (0 for no truncation)
:param pad: in seconds, will be added at the start and end before spectral transforms
:param get_raw: set to True to return raw computed spectrograms (e.g. for visualization)
:return:
"""
filename_midi = filename.rsplit('.')[0] + '.mid'
dname = filename.replace('/', '_').replace('\\', '_')
# Load files
ipad = int(pad * 44100)
audio_pad = (ipad, ipad
) # add one blank second at the beginning and at the end
if truncate > 0:
audio = AudioFile(filename,
truncate=int(truncate * 44100),
pad=audio_pad)
else:
audio = AudioFile(filename, pad=audio_pad)
mid = MidiFile(filename_midi)
step = 0.02 # seconds
latency = int(round(latency / step, 0))
# Compute spectrograms
spectrograms = ComputeSpectrograms(audio, step=step)
# Compute filtered spectrograms
melgrams = ComputeMelLayers(spectrograms, step, audio.Fs, latency)
# Build the input tensor
cnn_window = 15
tensor_mel = BuildTensor(melgrams[:, 2], cnn_window)
# Compute CQT
FreqAxisLog, time, cqgram = ComputeCqt(audio,
200.,
4000.,
step,
latency,
r=3)
tensor_cqt = BuildTensor([
cqgram,
], cnn_window)
# Global data length
max_len = min(tensor_mel.shape[0], tensor_cqt.shape[0])
# Compute output labels
notes = mid.getNotes(timidity)
notes_onset = np.array(notes)[:, 0] # get only the note timing
notes_value = np.array(notes, dtype=np.int)[:, 1] # get only the key value
onset_labels = np.zeros(max_len)
onset_caracs = np.zeros((max_len, 5))
onset_caracs[:, 2] = np.arange(max_len)
note_low = 21 # lowest midi note on a keyboard
note_high = 108 # highest midi note on a keyboard
notes_labels = np.zeros((max_len, note_high - note_low + 1))
notes_caracs = np.zeros((max_len, note_high - note_low + 1))
for i in range(len(notes_onset)):
t_win = int(np.floor(
(notes_onset[i] + audio_pad[0] / audio.Fs) / step))
if t_win >= len(onset_labels):
break
if t_win >= 0:
onset_labels[t_win] = 1
onset_caracs[t_win][0] += 1 # nb_notes
onset_caracs[t_win][1] = max(onset_caracs[t_win][1],
notes[i][2]) # max volume
if t_win + 1 < len(onset_labels):
onset_caracs[t_win + 1:, 2] -= onset_caracs[t_win + 1][
2] # nb of blank windows since the last onset
n = notes_value[i] - note_low
notes_labels[t_win][n] = 1
notes_caracs[t_win][n] = notes[i][2] # volume
counter = 0
for i in range(len(onset_labels) - 1, -1, -1):
onset_caracs[i][3] = counter
if onset_labels[i] == 1:
counter = 0
else:
counter += 1
onset_caracs[:, 4] = np.minimum(onset_caracs[:, 2], onset_caracs[:, 3])
# Extract useful CQT
select = [i for i in range(max_len) if onset_labels[i] > 0]
tensor_cqt_select = np.take(tensor_cqt, select, axis=0)
notes_labels_select = np.take(notes_labels, select, axis=0)
notes_caracs_select = np.take(notes_caracs, select, axis=0)
if not get_raw:
return (tensor_mel[:max_len, ...], tensor_cqt_select, onset_labels,
onset_caracs, notes_labels_select, notes_caracs_select, dname)
else:
return (melgrams, tensor_mel, onset_labels, cqgram, tensor_cqt, time,
FreqAxisLog, max_len, step)
