def test_spectrogram_db_magnituds_should_be_in_proper_range():
frame_size = 4096
hop_size = 4096
audio_file = os.path.join(DATA_DIR, 'she_brings_to_me.wav')
signal_frames = SignalFrames(audio_file,
frame_size,
hop_size,
mono_mix=True)
w = create_window(frame_size)
X = stft_spectrogram(signal_frames.frames, w, magnitudes='power_db')
assert np.all(X >= -120), 'min value: %f should be >= -120' % X.min()
assert np.all(X <= 0), 'max value: %f should be <= 0' % X.max()
def prepare_chomagram_and_labels(album, song_title, block_size, hop_size,
bin_range, bin_division):
song = 'The_Beatles/' + album + '/' + song_title
data_dir = '../data/beatles'
audio_file = data_dir + '/audio-cd/' + song + '.wav'
chord_file = data_dir + '/chordlab/' + song + '.lab.pcs.tsv'
audio_file, chord_file
# ## Load audio
print('loading audio:', audio_file)
x, fs = load_wav(audio_file)
print('sampling rate:', fs, 'Hz')
print('number of samples:', len(x))
print('duration in audio:', len(x) / fs, 'sec')
# ## Load chords
print('loading chords:', chord_file)
chords = pd.read_csv(chord_file, sep='\t')
print('shape:', chords.shape)
print('duration in chords:', chords['end'].iloc[-1])
pcs_cols = [
'C', 'Db', 'D', 'Eb', 'E', 'F', 'Gb', 'G', 'Ab', 'A', 'Bb', 'B'
]
label_cols = ['label', 'root', 'bass'] + pcs_cols
# ## Split audio to blocks
x_blocks, x_times = split_to_blocks(x, block_size, hop_size, fs)
print('blocks shape:', x_blocks.shape)
print('number of blocks:', len(x_blocks))
# start times for each block
print('last block starts at:', x_times[-1], 'sec')
# ## Mapping of chords to blocks
def chords_to_blocks(chords, block_center_times):
chord_ix = 0
for t in block_center_times:
yield chords.iloc[i][pcs_cols]
def time_to_samples(time):
return np.round(time * fs)
chords['start_sample'] = time_to_samples(chords['start'])
chords['end_sample'] = time_to_samples(chords['end'])
df_blocks = pd.DataFrame(
{'start': time_to_samples(x_times).astype(np.int64)})
df_blocks['end'] = df_blocks['start'] + block_size
label_dict = chords[label_cols].drop_duplicates().set_index('label')
df_labels = chords[['start_sample', 'end_sample', 'label']].copy()
df_labels.rename(columns={
'start_sample': 'start',
'end_sample': 'end'
},
inplace=True)
df_labelled_blocks = block_labels(df_blocks, df_labels)
df_block_pcs = df_labelled_blocks[['label']].join(
label_dict, on='label')[['label'] + pcs_cols]
assert len(df_block_pcs) == len(df_blocks)
block_labels_file = '{}/chord-pcs/{}_{}/{}.pcs'.format(
data_dir, block_size, hop_size, song)
print('block labels file:', block_labels_file)
os.makedirs(os.path.dirname(block_labels_file), exist_ok=True)
df_block_pcs.to_csv(block_labels_file, sep='\t', index=False)
# ## Chromagram features
w = create_window(block_size)
X_chromagram = chromagram(x_blocks,
w,
fs,
to_log=True,
bin_range=bin_range,
bin_division=bin_division)
chromagram_file = '{}/chromagram/block={}_hop={}_bins={},{}_div={}/{}.npz'.format(
data_dir, block_size, hop_size, bin_range[0], bin_range[1],
bin_division, song)
print('chomagram file:', chromagram_file)
os.makedirs(os.path.dirname(chromagram_file), exist_ok=True)
np.savez_compressed(chromagram_file, X=X_chromagram, times=x_times)
# features = data['X']
# times = data['times']
### Chord labels
# df_labels = pd.read_csv(labels_file, sep='\t')
# labels_pcs = df_labels[df_labels.columns[1:]].as_matrix()
block_size = 4096
hop_size = 2048
print('loading audio:', audio_file)
x, fs = load_wav(audio_file)
print('splitting audio to blocks')
x_blocks, times = split_to_blocks(x, block_size, hop_size)
w = create_window(block_size)
print('computing chromagram')
X_chromagram = chromagram(x_blocks, w, fs, to_log=True)
features = X_chromagram
## Data preprocessing
### Features
print('scaling the input features')
# scaler = MinMaxScaler()
# X = scaler.fit_transform(features).astype('float32')
# TODO: there's a bug: should be + 120 on both places!!!
X = (features.astype('float32') - 120) / (features.shape[1] - 120)
# reshape for 1D convolution
# features = data['X']
# times = data['times']
### Chord labels
# df_labels = pd.read_csv(labels_file, sep='\t')
# labels_pcs = df_labels[df_labels.columns[1:]].as_matrix()
block_size = 4096
hop_size = 2048
print('loading audio:', audio_file)
print('splitting audio to blocks')
signal_frames = SignalFrames(audio_file, frame_size=block_size, hop_size=hop_size)
x_blocks, x_times, fs = signal_frames.frames, signal_frames.start_times, signal_frames.sample_rate
w = create_window(block_size)
print('computing chromagram')
X_chromagram = chromagram(x_blocks, w, fs, to_log=True)
features = X_chromagram
## Data preprocessing
### Features
print('scaling the input features')
# scaler = MinMaxScaler()
# X = scaler.fit_transform(features).astype('float32')
# TODO: there's a bug: should be + 120 on both places!!!
X = (features.astype('float32') - 120) / (features.shape[1] - 120)
# reshape for 1D convolution
def assert_ok(size):
w = create_window(size)
assert np.allclose(energy(w), len(w))
assert np.allclose(mean_power(w), 1.0)