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)
1) * np.random.randn(n_fft - 1)
# Store the errors in the computed fs ratio
ratio_errors = np.zeros(len(timit_files))
# Store the RMS of the residual, before and after correction
residual_rms = np.zeros(len(timit_files))
original_rms = np.zeros(len(timit_files))
# Process each TIMIT file once
for n, filename in enumerate(timit_files):
a, fs = librosa.load(timit_files[n], sr=None)
# First, filter it
a_corrupted = np.convolve(a, random_filters[n])
# Then, skew it
a_corrupted = librosa.resample(a_corrupted, 1, fs_ratio_true[n],
'sinc_best')
# Zero pad so we can compute the residual RMS now...
a, a_corrupted = estimate.pad(a, a_corrupted)
original_rms[n] = np.sqrt(np.mean((a - a_corrupted)**2))
# Estimate the skew factor
estimated_ratio = estimate.get_best_fs_ratio(a_corrupted, a, .02, 400,
int(.05 * a.shape[0]),
int(.1 * a.shape[0]))
# Compute the relative error of this estimate
ratio_errors[n] = np.abs(fs_ratio_true[n] - estimated_ratio) / .04
# Reverse the skew
a_corrupted = librosa.resample(a_corrupted, 1, 1.0 / estimated_ratio)
a, a_corrupted = estimate.pad(a, a_corrupted)
# Compute STFTs
A_corrupted = librosa.stft(a_corrupted, n_fft=n_fft, hop_length=n_fft / 4)
A = librosa.stft(a, n_fft=256, hop_length=n_fft / 4)
A = np.array(A, dtype=np.complex128)
A_corrupted = np.array(A_corrupted, dtype=np.complex128)
# Then set h[n] = exp[n]*Normal(0, 1) for n > 0
random_filters[n, 1:] += np.exp( -np.arange( n_fft - 1 ) + 1 )*np.random.randn(n_fft - 1)
# Store the errors in the computed fs ratio
ratio_errors = np.zeros( len( timit_files ) )
# Store the RMS of the residual, before and after correction
residual_rms = np.zeros( len( timit_files ) )
original_rms = np.zeros( len( timit_files ) )
# Process each TIMIT file once
for n, filename in enumerate( timit_files ):
a, fs = librosa.load( timit_files[n], sr=None )
# First, filter it
a_corrupted = np.convolve( a, random_filters[n] )
# Then, skew it
a_corrupted = librosa.resample( a_corrupted, 1, fs_ratio_true[n], 'sinc_best' )
# Zero pad so we can compute the residual RMS now...
a, a_corrupted = estimate.pad( a, a_corrupted )
original_rms[n] = np.sqrt( np.mean( (a - a_corrupted)**2 ) )
# Estimate the skew factor
estimated_ratio = estimate.get_best_fs_ratio(a_corrupted, a, .02, 400, int(.05*a.shape[0]), int(.1*a.shape[0]))
# Compute the relative error of this estimate
ratio_errors[n] = np.abs( fs_ratio_true[n] - estimated_ratio)/.04
# Reverse the skew
a_corrupted = librosa.resample( a_corrupted, 1, 1.0/estimated_ratio )
a, a_corrupted = estimate.pad( a, a_corrupted )
# Compute STFTs
A_corrupted = librosa.stft( a_corrupted, n_fft=n_fft, hop_length=n_fft/4 )
A = librosa.stft( a, n_fft=256, hop_length=n_fft/4 )
A = np.array(A, dtype=np.complex128)
A_corrupted = np.array(A_corrupted, dtype=np.complex128)
# Estimate reverse filter
H = estimate.best_filter_coefficients( A, A_corrupted )
