def run(self):
# Initialize signals
self.percentDoneSignal.emit(0)
percent_scale = 1000.0 / 5
self.doneSignal.emit(0)
self.statusSignal.emit("")
# Load in audio data
self.statusSignal.emit("Loading {}".format(
os.path.split(self.mix_file)[1]))
mix, self.fs = librosa.load(self.mix_file, sr=None)
self.percentDoneSignal.emit(1 * percent_scale)
self.statusSignal.emit("Loading {}".format(
os.path.split(self.source_file)[1]))
source, self.fs = librosa.load(self.source_file, sr=self.fs)
self.percentDoneSignal.emit(2 * percent_scale)
# Fix any gross timing offset
self.statusSignal.emit("Aligning...")
mix, source = estimate.align(mix, source, self.fs)
self.percentDoneSignal.emit(3 * percent_scale)
self.statusSignal.emit("Subtracting...")
source = estimate.reverse_channel(mix, source)
mix, source = estimate.pad(mix, source)
self.percentDoneSignal.emit(4 * percent_scale)
self.statusSignal.emit("Enhancing...")
self.subtracted = estimate.wiener_enhance(mix - source, source,
self.wiener_threshold)
self.percentDoneSignal.emit(5 * percent_scale)
self.doneSignal.emit(1)
def run(self):
# Initialize signals
self.percentDoneSignal.emit(0)
percent_scale = 1000.0 / 5
self.doneSignal.emit(0)
self.statusSignal.emit("")
# Load in audio data
self.statusSignal.emit("Loading {}".format(os.path.split(self.mix_file)[1]))
mix, self.fs = librosa.load(self.mix_file, sr=None)
self.percentDoneSignal.emit(1 * percent_scale)
self.statusSignal.emit("Loading {}".format(os.path.split(self.source_file)[1]))
source, self.fs = librosa.load(self.source_file, sr=self.fs)
self.percentDoneSignal.emit(2 * percent_scale)
# Fix any gross timing offset
self.statusSignal.emit("Aligning...")
mix, source = estimate.align(mix, source, self.fs)
self.percentDoneSignal.emit(3 * percent_scale)
self.statusSignal.emit("Subtracting...")
source = estimate.reverse_channel(mix, source)
mix, source = estimate.pad(mix, source)
self.percentDoneSignal.emit(4 * percent_scale)
self.statusSignal.emit("Enhancing...")
self.subtracted = estimate.wiener_enhance(mix - source, source, self.wiener_threshold)
self.percentDoneSignal.emit(5 * percent_scale)
self.doneSignal.emit(1)
]
for subdirectory in [d for d in subdirectories if os.path.isdir(d)]:
print 'Processing file {}'.format(subdirectory)
# Load in mixture and corrupted source
mix, fs = librosa.load(os.path.join(subdirectory, 'M.wav'), sr=None)
source, fs = librosa.load(os.path.join(subdirectory, 'C.wav'), sr=fs)
# Align the source to the mixture
mix, source_aligned = estimate.align(mix, source, fs, max_global_offset=0)
# Estimate the filter
source_filtered = estimate.reverse_channel(mix, source_aligned)
# Write out aligned version
librosa.output.write_wav(os.path.join(subdirectory, 'C-filtered.wav'),
source_filtered, fs)
mix, source_filtered = estimate.pad(mix, source_filtered)
# Wiener filter the approximate separation
enhanced = estimate.wiener_enhance(mix - source_filtered, source_aligned,
6)
# Write out approximation of the true source
librosa.output.write_wav(os.path.join(subdirectory, 'S-approx.wav'),
enhanced, fs)
# <markdowncell>
# ## Step 2: Compute SDRs
# As above, we compute the SDR of our approximated source against the true source. We don't compare it against the true source with channel/timing distortion approximately removed because the vinyl distortion is much greater than for CDs.
# <codecell>
# Path to .wav files ripped from vinyl...
data_directory = '../Dataset/Vinyl/'
# Subdirectories 1-7 have C.wav an a cappella, 8-14 have instrumental
subdirectories = [
# Each example has its own folder, with M.wav (mix, CD), C.wav (vinyl corrupted source), and S.wav (true source, CD)
subdirectories = [os.path.join(data_directory, d) for d in os.listdir(data_directory)]
for subdirectory in [d for d in subdirectories if os.path.isdir(d)]:
print 'Processing file {}'.format(subdirectory)
# Load in mixture and corrupted source
mix, fs = librosa.load(os.path.join( subdirectory, 'M.wav' ), sr=None)
source, fs = librosa.load(os.path.join( subdirectory, 'C.wav' ), sr=fs)
# Align the source to the mixture
mix, source_aligned = estimate.align(mix, source, fs, max_global_offset=0)
# Estimate the filter
source_filtered = estimate.reverse_channel(mix, source_aligned)
# Write out aligned version
librosa.output.write_wav(os.path.join(subdirectory, 'C-filtered.wav'), source_filtered, fs)
mix, source_filtered = estimate.pad(mix, source_filtered)
# Wiener filter the approximate separation
enhanced = estimate.wiener_enhance( mix - source_filtered, source_aligned, 6 )
# Write out approximation of the true source
librosa.output.write_wav(os.path.join(subdirectory, 'S-approx.wav'), enhanced, fs)
# <markdowncell>
# ## Step 2: Compute SDRs
# As above, we compute the SDR of our approximated source against the true source. We don't compare it against the true source with channel/timing distortion approximately removed because the vinyl distortion is much greater than for CDs.
# <codecell>
# Path to .wav files ripped from vinyl...
data_directory = '../Dataset/Vinyl/'
# Subdirectories 1-7 have C.wav an a cappella, 8-14 have instrumental
subdirectories = [os.path.join(data_directory, d) for d in os.listdir(data_directory)]
# Dict to store the results, each entry is an array of length 7
