def test_nmf_mfcc_initialization(self):
# Test initializing the MFCC range
# Set up the max of the MFCC range by only using an int
nmf_mfcc = nussl.NMF_MFCC(self.signal_mono,
num_sources=self.n_src,
num_templates=6,
num_iterations=5,
mfcc_range=5)
assert nmf_mfcc.mfcc_start == 1
assert nmf_mfcc.mfcc_end == 5
nmf_mfcc.run()
nmf_mfcc.make_audio_signals()
# Set up the MFCC range by using a list [min, max]
nmf_mfcc = nussl.NMF_MFCC(self.signal_mono,
num_sources=self.n_src,
num_templates=6,
num_iterations=5,
mfcc_range=[3, 15])
assert nmf_mfcc.mfcc_start == 3
assert nmf_mfcc.mfcc_end == 15
# Set up the MFCC range by using a tuple (min, max)
nmf_mfcc = nussl.NMF_MFCC(self.signal_mono,
num_sources=self.n_src,
num_templates=6,
num_iterations=5,
mfcc_range=(2, 14))
assert nmf_mfcc.mfcc_start == 2
assert nmf_mfcc.mfcc_end == 14
def test_random_seed_initialization(self):
# Different random_seed and random_state, check if set separately
kmean_kwargs = {'random_state': 0}
nmf_mfcc = nussl.NMF_MFCC(self.signal_mono,
num_sources=self.n_src,
num_templates=6,
num_iterations=5,
random_seed=1,
kmeans_kwargs=kmean_kwargs)
assert nmf_mfcc.clusterer.random_state == 0
assert nmf_mfcc.random_seed == 1
# No random_seed, but random_state initialized
kmean_kwargs = {'random_state': 0}
nmf_mfcc = nussl.NMF_MFCC(self.signal_mono,
num_sources=2,
num_templates=6,
num_iterations=5,
kmeans_kwargs=kmean_kwargs)
assert nmf_mfcc.clusterer.random_state == 0
assert nmf_mfcc.random_seed is None
# No random_state, but random_seed initialized
nmf_mfcc = nussl.NMF_MFCC(self.signal_mono,
num_sources=self.n_src,
num_templates=6,
num_iterations=5,
random_seed=0)
assert nmf_mfcc.clusterer.random_state == nmf_mfcc.random_seed == 0
def main():
"""
A simple example of Non-Negative Matrix Factorization (NMF) with k-Means MFCC clustering in nussl.
See nussl documentation for more information about using NMF_MFCC.
Returns:
"""
# Load input file
input_file_name = os.path.join('..', 'input',
'piano_and_synth_arp_chord_mono.wav')
signal = nussl.AudioSignal(path_to_input_file=input_file_name)
# make a directory to store output if needed
if not os.path.exists(os.path.join('..', 'Output/')):
os.mkdir(os.path.join('..', 'Output/'))
# Set up NMMF MFCC
nmf_mfcc = nussl.NMF_MFCC(signal,
num_sources=2,
num_templates=6,
num_iterations=10,
random_seed=0,
distance_measure=nussl.TransformerNMF.EUCLIDEAN)
# and run
nmf_mfcc.run()
sources = nmf_mfcc.make_audio_signals()
for i, source in enumerate(sources):
output_file_name = '{}_{}.wav'.format(
os.path.splitext(signal.file_name)[0], i)
source.write_audio_to_file(output_file_name)
def main():
"""
Freezes essential values from NMF MFCC in its current implementation for benchmarking
See test_benchmark_nmf_mfcc() in test_nmf_mfcc.py for usage.
Run with top level ('nussl/') as working directory
"""
metadata = {}
path_to_input_file = os.path.join('input',
'piano_and_synth_arp_chord_mono.wav')
metadata['input_file'] = path_to_input_file
signal = nussl.AudioSignal(path_to_input_file)
# Set random seed in NMF and KMeans to 0
params = {
'num_sources': 2,
'num_templates': 6,
'distance_measure': nussl.transformers.TransformerNMF.EUCLIDEAN,
'num_iterations': 10,
'random_seed': 0
}
metadata['params'] = params
nmf_mfcc = nussl.NMF_MFCC(signal, **params)
if DEBUG:
output_folder = os.path.join('tests', 'nmf_mfcc_reference', 'scratch')
if not os.path.exists(output_folder):
os.mkdir(output_folder)
else:
output_folder = os.path.join('tests', 'nmf_mfcc_reference',
'nmf_mfcc_benchmark_files')
nmf_mfcc.run()
np.save(os.path.join(output_folder, 'benchmark_labeled_templates'),
nmf_mfcc.labeled_templates)
np.save(os.path.join(output_folder, 'benchmark_masks'),
nmf_mfcc.result_masks)
nmf_mfcc.make_audio_signals()
# Make sure the paths are empty
for source in nmf_mfcc.sources:
source.path_to_input_file = ''
np.save(os.path.join(output_folder, 'benchmark_sources'), nmf_mfcc.sources)
metadata['save_time'] = time.asctime()
metadata['nussl_version'] = nussl.version
metadata['made_by'] = platform.uname()[1]
json.dump(
metadata,
open(os.path.join(output_folder, 'nmf_mfcc_benchmark_metadata.json'),
'w'))
def test_nmf_mfcc(self):
path = os.path.join('..', 'Input',
'piano_and_synth_arp_chord_mono.wav')
a = nussl.AudioSignal(path)
n = nussl.NMF_MFCC(a, num_sources=2)
n()
j = n.to_json()
f = nussl.NMF_MFCC.from_json(j)
worked = n == f
return worked
def test_piano_and_synth_stereo(self):
# Set up NMMF MFCC
nmf_mfcc = nussl.NMF_MFCC(self.signal_stereo, num_sources=self.n_src,
num_templates=6, num_iterations=15,
random_seed=0)
# and run
nmf_mfcc.run()
estimated_sources = nmf_mfcc.make_audio_signals()
assert len(estimated_sources) == self.n_src
for source in estimated_sources:
assert source.is_stereo
def test_piano_and_synth(self):
# Set up NMMF MFCC
nmf_mfcc = nussl.NMF_MFCC(self.signal_mono, num_sources=self.n_src * 2,
num_templates=6, num_iterations=10,
random_seed=0)
# and run
computed_masks = nmf_mfcc.run()
estimated_sources = nmf_mfcc.make_audio_signals()
assert len(estimated_sources) == self.n_src
for source in estimated_sources:
assert source.is_mono
assert self.signal_mono == estimated_sources[0] + estimated_sources[1]
def test_piano_and_synth_stereo_to_mono(self):
# Set up NMMF MFCC and convert input audio signal to mono
nmf_mfcc = nussl.NMF_MFCC(self.signal_stereo, num_sources=self.n_src,
num_templates=6, num_iterations=15,
random_seed=0, distance_measure='euclidean',
to_mono=True)
assert nmf_mfcc.audio_signal.is_mono
# and run
nmf_mfcc.run()
estimated_sources = nmf_mfcc.make_audio_signals()
assert len(estimated_sources) == self.n_src
for source in estimated_sources:
assert source.is_mono
def test_2_sin_waves(self):
sr = nussl.DEFAULT_SAMPLE_RATE
dur = 3 # seconds
length = dur * sr
sine_wave_1 = np.sin(np.linspace(0, 440 * 2 * np.pi, length))
sine_wave_2 = np.sin(np.linspace(0, 440 * 5 * 2 * np.pi, length))
signal = sine_wave_1 + sine_wave_2
test_audio = nussl.AudioSignal()
test_audio.load_audio_from_array(signal)
# Set up NMMF MFCC
nmf_mfcc = nussl.NMF_MFCC(test_audio, num_sources=self.n_src,
num_templates=2, num_iterations=15)
# and run
nmf_mfcc.run()
nmf_mfcc.make_audio_signals()
def audio_decomp(video=None, audio=None, OutputPath=None, file=None):
signal = nussl.AudioSignal(audio_data_array=audio[0, :], sample_rate=44100)
nmf_mfcc = nussl.NMF_MFCC(signal,
num_sources=2,
num_templates=25,
distance_measure="euclidean",
random_seed=0)
nmf_mfcc.run()
sources = nmf_mfcc.make_audio_signals()
source_list = []
for i, source in enumerate(sources):
outputfile = os.path.join(OutputPath,
file + '_seg' + str(i + 1) + '.wav')
source.write_audio_to_file(outputfile)
source_list.append(source.audio_data)
return [
os.path.join(OutputPath, file + "_seg1.wav"),
os.path.join(OutputPath, file + "_seg2.wav")
], source_list
def test_benchmark_nmf_mfcc(self):
metadata_file = os.path.join('tests', 'nmf_mfcc_reference',
'nmf_mfcc_benchmark_files',
'nmf_mfcc_benchmark_metadata.json')
metadata = json.load(open(metadata_file, 'r'))
print('Reading benchmark files from {}, version {}, made by {}'.format(
metadata['save_time'], metadata['nussl_version'],
metadata['made_by']))
# Load the audio
signal = nussl.AudioSignal(metadata['input_file'])
# Set up NMMF MFCC
params = metadata['params'] # Load params from metadata
nmf_mfcc = nussl.NMF_MFCC(signal, **params)
# run
nmf_mfcc.run()
benchmark_labeled_templates = np.load(
os.path.join('tests', 'nmf_mfcc_reference',
'nmf_mfcc_benchmark_files',
'benchmark_labeled_templates.npy'))
benchmark_masks = np.load(
os.path.join('tests', 'nmf_mfcc_reference',
'nmf_mfcc_benchmark_files', 'benchmark_masks.npy'))
nmf_mfcc.make_audio_signals()
# Set the paths to empty so we can compare
for source in nmf_mfcc.sources:
source.path_to_input_file = ''
benchmark_sources = np.load(
os.path.join('tests', 'nmf_mfcc_reference',
'nmf_mfcc_benchmark_files', 'benchmark_sources.npy'))
assert np.all(benchmark_sources == nmf_mfcc.sources)
assert np.all(
benchmark_labeled_templates == nmf_mfcc.labeled_templates)
assert np.all(benchmark_masks == nmf_mfcc.result_masks)