def main():
OutputPath = 'result_audio'
namelist = os.listdir('testimage')
for name in namelist:
signal1 = nussl.AudioSignal(path_to_input_file='gt_audio/' + name +
'_gt1.wav')
signal2 = nussl.AudioSignal(path_to_input_file='gt_audio/' + name +
'_gt2.wav')
signal = nussl.AudioSignal(path_to_input_file='gt_audio/' + name +
'.wav')
audio = signal.audio_data
m = audio_decomp(audio=audio, OutputPath=OutputPath, file=name)
signal3 = nussl.AudioSignal(path_to_input_file='result_audio/' + name +
'_seg1.wav')
signal4 = nussl.AudioSignal(path_to_input_file='result_audio/' + name +
'_seg2.wav')
ref_sources = np.zeros([2, len(audio[0, :])])
est_sources = np.zeros([2, len(audio[0, :])])
ref_sources[0, :] = sum(signal1.audio_data)
ref_sources[1, :] = sum(signal2.audio_data)
est_sources[0, :] = sum(signal3.audio_data)
est_sources[1, :] = sum(signal4.audio_data)
result = separation.bss_eval_sources(ref_sources,
est_sources,
compute_permutation=True)
print result[0]
def test_write_to_file_path1(self):
a = nussl.AudioSignal(self.input_path1)
a.write_audio_to_file(self.out_path1)
b = nussl.AudioSignal(self.out_path1)
assert (a.sample_rate == b.sample_rate)
assert (np.allclose(a.audio_data, b.audio_data))
def main():
win_type = nussl.constants.WINDOW_HANN
sample_rate = nussl.DEFAULT_SAMPLE_RATE
# Plot a simple sine wave at 5kHz
dt = 1.0 / float(sample_rate)
dur = 10.0 # sec
freq = 5000 # Hz
x = np.arange(0.0, dur, dt)
x = np.sin(2 * np.pi * freq * x)
signal = nussl.AudioSignal(audio_data_array=x)
nussl.plot_stft(signal.get_channel(1), '../Output/sine_wave5000Hz.png', window_type=win_type)
# Make a FM signal and plot it.
lfo_freq = 3
lfo_amp = 800
freq2 = 15000
x2 = np.arange(0.0, dur, dt)
modulator = lfo_amp * np.sin(2 * np.pi * lfo_freq * x2)
x2 = np.sin(2 * np.pi * freq2 * x2 + modulator)
x2 += x
signal2 = nussl.AudioSignal(audio_data_array=x2)
nussl.plot_stft(signal2.get_channel(1), '../Output/fm_wave.png', window_type=win_type)
# Plot from a file
path = '../Input/police_noisy.wav'
a = nussl.AudioSignal(path)
nussl.plot_stft(a.get_channel(1), '../Output/police_noisy.png', freq_max=18000)
# Plot through audio_signal
a.plot_spectrogram()
def test_load(benchmark_audio):
# Load from file
a = nussl.AudioSignal(benchmark_audio['K0140.wav'])
b = nussl.AudioSignal()
b.load_audio_from_file(benchmark_audio['K0140.wav'])
assert (np.array_equal(a.audio_data, b.audio_data))
assert (a.sample_rate == b.sample_rate)
# Load from array
ref_sr, ref_data = wav.read(benchmark_audio['K0140.wav'])
c = nussl.AudioSignal(audio_data_array=ref_data, sample_rate=ref_sr)
pytest.raises(Exception, nussl.AudioSignal, benchmark_audio['K0140.wav'],
ref_data)
pytest.raises(Exception,
nussl.AudioSignal,
path_to_input_file=benchmark_audio['K0140.wav'],
audio_data_array=ref_data)
d = nussl.AudioSignal()
d.load_audio_from_array(ref_data, ref_sr)
assert (np.array_equal(c.audio_data, d.audio_data))
assert (c.sample_rate == d.sample_rate)
assert (b.sample_rate == c.sample_rate)
assert (np.array_equal(b.audio_data, c.audio_data))
def test_stft_istft_defaults(benchmark_audio, atol=stft_tol):
dummy = nussl.AudioSignal()
pytest.raises(AudioSignalException, dummy.stft)
a = nussl.AudioSignal(audio_data_array=sine_wave)
pytest.raises(AudioSignalException, a.istft)
a.stft()
a.istft()
a = nussl.AudioSignal(audio_data_array=sine_wave)
a.stft()
calc_sine = a.istft(overwrite=False)
assert np.allclose(a.audio_data, calc_sine, atol=stft_tol)
# also load another object with stft_data
b = nussl.AudioSignal(stft=a.stft(), sample_rate=a.sample_rate)
b.istft()
min_length = min(b.audio_data.shape[1], a.audio_data.shape[1])
assert np.allclose(a.audio_data[:, :min_length],
b.audio_data[:, :min_length],
atol=stft_tol)
for key, path in benchmark_audio.items():
a = nussl.AudioSignal(path)
a.stft()
recon = a.istft(overwrite=False)
assert np.allclose(a.audio_data, recon, atol=stft_tol)
def test_to_mono(self):
"""
Test functionality and correctness of AudioSignal.to_mono() function.
Returns:
"""
num_samples = nussl.DEFAULT_SAMPLE_RATE # 1 second
sin1 = np.sin(np.linspace(0, 100 * 2 * np.pi,
num_samples)) # Freq = 100 Hz
sig1 = nussl.AudioSignal(audio_data_array=sin1)
assert (sig1.num_channels == 1)
sig1.to_mono(overwrite=True)
assert (sig1.num_channels == 1)
sin2 = -1 * sin1
sines = np.vstack((sin1, sin2))
sig2 = nussl.AudioSignal(audio_data_array=sines)
assert (sig2.num_channels == 2)
sig2.to_mono()
assert (sig2.num_channels == 2)
sig2.to_mono(overwrite=False)
assert (sig2.num_channels == 2)
sig2.to_mono(overwrite=True)
assert (sig2.num_channels == 1)
assert (np.allclose([0.0] * len(sig2), sig2.audio_data))
def test_write_sample_rate(self):
a = nussl.AudioSignal(self.audio_input1)
sample_rate = a.sample_rate // 2
a.write_audio_to_file(self.audio_output, sample_rate=sample_rate)
b = nussl.AudioSignal(self.audio_output)
assert (b.sample_rate == sample_rate)
def test_properties(benchmark_audio):
a = nussl.AudioSignal()
assert a.signal_duration is None
assert a.signal_length is None
assert a._signal_length is None
assert a.num_channels is None
assert a.time_vector is None
assert a.file_name is None
assert not a.has_data
a = nussl.AudioSignal(benchmark_audio['K0140.wav'])
assert a.has_data
assert a.file_name == 'K0140.wav'
assert len(a.time_vector) == a.audio_data.shape[-1]
pytest.raises(AudioSignalException, lambda x: x.stft_length, a)
pytest.raises(AudioSignalException, lambda x: x.time_bins_vector, a)
pytest.raises(AudioSignalException, lambda x: x.freq_vector, a)
a.stft()
assert a.stft_length == a.stft_data.shape[1]
assert len(a.time_bins_vector) == a.stft_length
assert len(a.freq_vector) == a.stft_data.shape[0]
assert len(a.freq_vector) == a.stft_data.shape[0]
a.audio_data = None
assert a.has_data
def test_write_sample_rate(self):
a = nussl.AudioSignal(self.input_path1)
sample_rate = a.sample_rate // 2
a.write_audio_to_file(self.out_path1, sample_rate=sample_rate)
b = nussl.AudioSignal(self.out_path1)
assert (b.sample_rate == sample_rate)
def test_simple(self):
drums_path = os.path.join('Input', 'src1.wav')
flute_path = os.path.join('Input', 'src2.wav')
drums = nussl.AudioSignal(drums_path)
flute = nussl.AudioSignal(flute_path)
flute.truncate_samples(drums.signal_length)
gains = [1.0, 0.75, 0.5, 0.25, 0.0] # gain settings
drum_sigs = [
drums.make_copy_with_audio_data(drums.audio_data * g)
for g in gains
] # drums with different gains
mixtures = [d + flute for d in drum_sigs] # mix everything together
true_sources = [[flute, d] for d in drum_sigs]
repet_sim = nussl.RepetSim
repet_kwargs = {}
scores_sim = nussl.run_and_eval_prf(repet_sim, repet_kwargs, mixtures,
true_sources)
repet = nussl.Repet
scores_repet = nussl.run_and_eval_prf(repet, repet_kwargs, mixtures,
true_sources)
i = 0
def test_utils_audio_signals_to_musdb_track(musdb_tracks):
track = musdb_tracks[0]
mixture = nussl.AudioSignal(audio_data_array=track.audio,
sample_rate=track.rate)
mixture.stft()
stems = track.stems
true_sources = {}
fake_sources = {}
for k, v in sorted(track.sources.items(), key=lambda x: x[1].stem_id):
true_sources[k] = nussl.AudioSignal(audio_data_array=stems[v.stem_id],
sample_rate=track.rate)
mask_data = np.random.rand(*mixture.stft_data.shape)
soft_mask = SoftMask(mask_data)
_source = mixture.apply_mask(soft_mask)
_source.istft(truncate_to_length=mixture.signal_length)
fake_sources[k] = _source
separated_track = nussl.utils.audio_signals_to_musdb_track(
mixture, fake_sources, nussl.constants.STEM_TARGET_DICT)
reconstructed_track = nussl.utils.audio_signals_to_musdb_track(
mixture, true_sources, nussl.constants.STEM_TARGET_DICT)
assert np.allclose(track.stems, reconstructed_track.stems)
assert track.stems.shape == separated_track.stems.shape
def test_sr_on_load_from_array(self):
# Check that the passed in sample rate is being set in load_audio_from_array
a = nussl.AudioSignal(self.audio_input1)
sr, data = wav.read(self.audio_input1)
b = nussl.AudioSignal()
b.load_audio_from_array(data, sample_rate=sr)
assert (a.sample_rate == b.sample_rate)
assert (np.allclose(a.audio_data, b.audio_data))
def test_resample_on_load_from_file(self):
# Test resample right when loading from file vs resampling after loading
a = nussl.AudioSignal(self.audio_input1)
a.resample(48000)
b = nussl.AudioSignal()
b.load_audio_from_file(self.audio_input1, new_sample_rate=48000)
assert (a.sample_rate == b.sample_rate)
assert (np.allclose(a.audio_data, b.audio_data))
def test_write_to_file_array1(self):
sr, data = wav.read(self.input_path1)
a = nussl.AudioSignal(audio_data_array=data, sample_rate=sr)
a.write_audio_to_file(self.out_path1)
b = nussl.AudioSignal(self.out_path1)
assert (a.sample_rate == b.sample_rate)
assert (np.allclose(a.audio_data, b.audio_data))
def test_write_to_file_path2(self):
a = nussl.AudioSignal()
a.load_audio_from_file(self.audio_input1)
a.write_audio_to_file(self.audio_output)
b = nussl.AudioSignal(self.audio_output)
assert (a.sample_rate == b.sample_rate)
assert (np.allclose(a.audio_data, b.audio_data))
def main():
model_path = "/Users/ethanmanilow/Documents/School/Research/audio_representations/website/backend/models/data/models/deep_clustering_vocals_44k_long.model"
cutoff = -40
output_dir = os.path.join("output", "tdc_test")
# for song_path in dsd_folder:
# Load 'em all into memory
song_path = "/Users/ethanmanilow/Documents/School/Research/audio_representations/website/backend/scripts/test_files"
mix_path = os.path.join(song_path, "mix.wav")
vox_path = os.path.join(song_path, "vox.wav")
bk_path = os.path.join(song_path, "gtr.wav")
mix = nussl.AudioSignal(mix_path)
mix.to_mono(overwrite=True)
vox = nussl.AudioSignal(vox_path)
vox.to_mono(overwrite=True)
bak = nussl.AudioSignal(bk_path)
bak.to_mono(overwrite=True)
n_bins = 100
if RUN_DC:
gt_vox_mask, gt_bak_mask, sm = ct.mel_mask(
mix,
vox,
bak,
mix.sample_rate,
bg_mask_inverse=False,
silence_mask_cutoff=cutoff,
)
dc = nussl.DeepClustering(
mix,
model_path=model_path,
mask_type="binary",
do_mono=True,
return_mel_masks=True,
pca_before_clustering=False,
cutoff=cutoff,
)
dc_vox_mask, dc_bk_mask, binned, mel, scaled = ct.deep_clustering_mask(
dc)
dc_vox_mask, dc_bk_mask = dc_vox_mask.get_channel(
0), dc_bk_mask.get_channel(0)
else:
dc_vox_mask = np.load(os.path.join("pickles", "dc_vox_mask.npy"))
dc_bk_mask = np.load(os.path.join("pickles", "dc_bk_mask.npy"))
binned = np.load(os.path.join("pickles", "binned.npy"))
scaled = np.load(os.path.join("pickles", "scaled.npy"))
tdc_space = np.zeros((n_bins, dc_vox_mask.shape[0]))
for (i, j), tf_list in np.ndenumerate(binned):
for item in tf_list:
t, f = get_coordinate_from_TF_index(item, dc_vox_mask.shape[1])
tdc_space[j, t] += 1
plot_tdc(tdc_space, mix.time_vector[-1], output_dir, "tdc_test3.png")
def test_write_to_file_array2(self):
sr, data = wav.read(self.audio_input1)
a = nussl.AudioSignal()
a.load_audio_from_array(data, sr)
a.write_audio_to_file(self.audio_output)
b = nussl.AudioSignal(self.audio_output)
assert (a.sample_rate == b.sample_rate)
assert (np.allclose(a.audio_data, b.audio_data))
def test_resample(benchmark_audio):
# Check that sample rate property changes
for key, path in benchmark_audio.items():
a = nussl.AudioSignal(path)
b = nussl.AudioSignal(path)
b.resample(a.sample_rate / 2)
assert (b.sample_rate == a.sample_rate / 2)
pytest.warns(UserWarning, a.resample, a.sample_rate)
def test_write_sample_rate(benchmark_audio):
for key, path in benchmark_audio.items():
with tempfile.NamedTemporaryFile(suffix='.wav', delete=True) as f:
a = nussl.AudioSignal(path)
sample_rate = a.sample_rate // 2
a.write_audio_to_file(f.name, sample_rate=sample_rate)
b = nussl.AudioSignal(f.name)
assert (b.sample_rate == sample_rate)
def test_multiple_duet(self):
benchmark_mask = self.benchmark_dict['benchmark_masks']
duet = nussl.Duet(self.signal, 3)
duet.run()
duet.audio_signal = nussl.AudioSignal(self.dev1_wdrums)
duet.run()
duet.audio_signal = nussl.AudioSignal(self.dev1_female3)
duet_masks = duet.run()
for i in range(len(duet_masks)):
assert np.array_equal(benchmark_mask[i].mask, duet_masks[i].mask)
def test_write_array_to_file(benchmark_audio):
for key, path in benchmark_audio.items():
with tempfile.NamedTemporaryFile(suffix='.wav', delete=True) as f:
sr, data = wav.read(path)
a = nussl.AudioSignal(audio_data_array=data, sample_rate=sr)
a.write_audio_to_file(f.name)
b = nussl.AudioSignal(f.name)
assert (a.sample_rate == b.sample_rate)
assert (np.allclose(a.audio_data, b.audio_data))
def test_resample_on_load_from_file(benchmark_audio):
# Test resample right when loading from file vs resampling after loading
path = [benchmark_audio[key] for key in benchmark_audio][0]
a = nussl.AudioSignal(path)
a.resample(48000)
b = nussl.AudioSignal()
b.load_audio_from_file(path, new_sample_rate=48000)
assert (a.sample_rate == b.sample_rate)
assert (np.allclose(a.audio_data, b.audio_data))
pytest.warns(UserWarning, nussl.AudioSignal, path, sample_rate=16000)
def setUp(self):
dur = 30
offset = 60
premade_mixture_path = os.path.join('..', 'input', 'mixture', 'mixture.wav')
vocals_path = os.path.join('..', 'input', 'mixture', 'vocals.wav')
drums_path = os.path.join('..', 'input', 'mixture', 'drums.wav')
self.premade_mixture = nussl.AudioSignal(premade_mixture_path, duration=dur, offset=offset)
self.vocals = nussl.AudioSignal(vocals_path, duration=dur, offset=offset)
self.drums = nussl.AudioSignal(drums_path, duration=dur, offset=offset)
self.new_mixture = self.vocals + self.drums
def test_write_to_file(benchmark_audio):
for key, path in benchmark_audio.items():
with tempfile.NamedTemporaryFile(suffix='.wav', delete=True) as f:
dummy = nussl.AudioSignal()
pytest.raises(AudioSignalException, dummy.write_audio_to_file,
f.name)
a = nussl.AudioSignal(path)
a.write_audio_to_file(f.name)
b = nussl.AudioSignal(f.name)
assert (a.sample_rate == b.sample_rate)
def setUp(self):
# If our working directory is not the top level dir
if os.path.basename(os.path.normpath(os.getcwd())) == 'tests':
os.chdir('..') # then go up one level
input_mono = os.path.join('input', 'piano_and_synth_arp_chord_mono.wav')
self.signal_mono = nussl.AudioSignal(input_mono)
input_stereo = os.path.join('input', 'piano_and_synth_arp_chord_stereo.wav')
self.signal_stereo = nussl.AudioSignal(input_stereo)
self.n_src = 2 # number of sources in both of these files
def test_stft_istft_simple2(self):
a = nussl.AudioSignal(audio_data_array=self.sine_wave)
a.stft(use_librosa=True)
calc_sine = a.istft(overwrite=False, use_librosa=True)
assert np.allclose(a.audio_data, calc_sine, atol=1e-3)
a = nussl.AudioSignal(audio_data_array=self.sine_wave)
a.stft(use_librosa=False)
calc_sine = a.istft(overwrite=False, use_librosa=False)
assert np.allclose(a.audio_data, calc_sine)
def _get_signal(with_stems=False, track='Track00001'):
babyslakh = os.path.expanduser('~/Documents/School/Research/babyslakh_16k')
track = os.path.join(babyslakh, track)
mix = nussl.AudioSignal(os.path.join(track, 'mix.wav'))
if not with_stems:
return mix
stem_names = os.listdir(os.path.join(babyslakh, track, 'stems'))
stems = {}
for s in stem_names:
if os.path.splitext(s)[1] == '.wav':
stem_path = os.path.join(babyslakh, track, 'stems', s)
stems[os.path.splitext(s)[0]] = nussl.AudioSignal(stem_path)
return mix, stems
def estimated_and_true_sources(musdb_tracks):
i = np.random.randint(len(musdb_tracks))
track = musdb_tracks[i]
mixture = nussl.AudioSignal(audio_data_array=track.audio,
sample_rate=track.rate)
mixture.stft()
stems = track.stems
oracle_sources = []
random_sources = []
true_sources = []
random_masks = []
oracle_masks = []
keys = []
for k, v in sorted(track.sources.items(), key=lambda x: x[1].stem_id):
true_sources.append(
nussl.AudioSignal(audio_data_array=stems[v.stem_id],
sample_rate=track.rate))
keys.append(k)
mask_data = np.random.rand(*mixture.stft_data.shape)
random_mask = SoftMask(mask_data)
random_source = mixture.apply_mask(random_mask)
random_source.istft(truncate_to_length=mixture.signal_length)
random_sources.append(random_source)
random_masks.append(random_mask)
source_stft = true_sources[-1].stft()
mask_data = ((np.abs(source_stft) + 1e-8) / (
np.maximum(np.abs(mixture.stft_data), np.abs(source_stft)) + 1e-8))
oracle_mask = SoftMask(mask_data)
oracle_source = mixture.apply_mask(oracle_mask)
oracle_source.istft(truncate_to_length=mixture.signal_length)
oracle_sources.append(oracle_source)
oracle_masks.append(oracle_mask)
yield {
'oracle': oracle_sources,
'random': random_sources,
'true': true_sources,
'keys': keys,
'oracle_masks': oracle_masks,
'random_masks': random_masks,
}
def mix_source_folder(toy_datasets):
wsj_sources = toy_datasets['babywsj_oW0F0H9.zip']
audio_files = glob.glob(f"{wsj_sources}/**/*.wav", recursive=True)
n_sources = 2
n_mixtures = 10
with tempfile.TemporaryDirectory() as tmp_dir:
_dir = tmp_dir if fix_dir is None else fix_dir
_dir = os.path.join(_dir, 'mix_source_folder')
for i in range(n_mixtures):
sources = []
for n in range(n_sources):
path = random.choice(audio_files)
source = nussl.AudioSignal(path)
sources.append(source)
min_length = min([s.signal_length for s in sources])
for n in range(n_sources):
output_path = os.path.join(_dir, f's{n}', f'{i}.wav')
os.makedirs(os.path.dirname(output_path), exist_ok=True)
sources[n].truncate_samples(min_length)
sources[n].write_audio_to_file(output_path)
mix = sum(sources)
output_path = os.path.join(_dir, 'mix', f'{i}.wav')
os.makedirs(os.path.dirname(output_path), exist_ok=True)
mix.write_audio_to_file(output_path)
yield _dir
def _get_channel_helper(signal, n_channels):
a = nussl.AudioSignal(audio_data_array=signal)
# Check that we are counting our channels correctly
assert a.num_channels == n_channels
# Check that we can get every channel with AudioSignal.get_channel()
for i, ch in enumerate(signal):
assert np.array_equal(a.get_channel(i), ch)
# Check that attempting to get higher channels raises exception
for i in range(n_channels, n_channels + 10):
pytest.raises(AudioSignalException, a.get_channel, i)
# Check that attempting to get lower channels raises exception
for i in range(-1, -11, -1):
pytest.raises(AudioSignalException, a.get_channel, i)
# Check that AudioSignal.get_channels() generator works
i = 0
for ch in a.get_channels():
assert np.array_equal(ch, signal[i, :])
new_signal = a.make_audio_signal_from_channel(i)
assert np.array_equal(ch, new_signal.audio_data[0])
i += 1
assert i == a.num_channels
