def incorrect_input_size(D, num_frames):
if D == 1:
x_local = x[:, 0]
else:
x_local = x[:, :D]
# parameters
block_size = 512
hop = block_size
# create STFT object
stft = STFT(block_size,
hop=hop,
channels=D,
transform=transform,
num_frames=num_frames)
try: # passing more frames than 'hop'
stft.analysis(x_local)
computed = False
except:
computed = True
return computed
def no_overlap_no_filter(D, num_frames=1, fixed_memory=False, streaming=True):
"""
D - number of channels
num_frames - how many frames to process, None will process one frame at
a time
fixed_memory - whether to enforce checks for size (real-time consideration)
streaming - whether or not to stitch between frames
"""
if D == 1:
x_local = x[:, 0]
else:
x_local = x[:, :D]
# parameters
block_size = 512 # make sure the FFT size is a power of 2
hop = block_size # no overlap
if not streaming:
num_samples = (num_frames - 1) * hop + block_size
x_local = x_local[:num_samples, ]
# Create the STFT object
if fixed_memory:
stft = STFT(
block_size,
hop=hop,
channels=D,
transform=transform,
num_frames=num_frames,
streaming=streaming,
)
else:
stft = STFT(block_size,
hop=hop,
channels=D,
transform=transform,
streaming=streaming)
# collect the processed blocks
processed_x = np.zeros(x_local.shape)
if streaming:
n = 0
hop_frames = hop * num_frames
# process the signals while full blocks are available
while x_local.shape[0] - n > hop_frames:
stft.analysis(x_local[n:n + hop_frames, ])
processed_x[n:n + hop_frames, ] = stft.synthesis()
n += hop_frames
else:
stft.analysis(x_local)
processed_x = stft.synthesis()
n = processed_x.shape[0]
error = np.max(np.abs(x_local[:n, ] - processed_x[:n, ]))
return error
def apply_spectral_sub(
noisy_signal, nfft=512, db_reduc=25, lookback=12, beta=30, alpha=1
):
"""
One-shot function to apply spectral subtraction approach.
Parameters
----------
noisy_signal : numpy array
Real signal in time domain.
nfft: int
FFT size. Length of gain filter, i.e. the number of frequency bins, is
given by ``nfft//2+1``.
db_reduc: float
Maximum reduction in dB for each bin.
lookback: int
How many frames to look back for the noise estimate.
beta: float
Overestimation factor to "push" the gain filter value (at each
frequency) closer to the dB reduction specified by ``db_reduc``.
alpha: float, optional
Exponent factor to modify transition behavior towards the dB reduction
specified by ``db_reduc``. Default is 1.
Returns
-------
numpy array
Enhanced/denoised signal.
"""
from pyroomacoustics import hann
from pyroomacoustics.transform import STFT
hop = nfft // 2
window = hann(nfft, flag="asymmetric", length="full")
stft = STFT(nfft, hop=hop, analysis_window=window, streaming=True)
scnr = SpectralSub(nfft, db_reduc, lookback, beta, alpha)
processed_audio = np.zeros(noisy_signal.shape)
n = 0
while noisy_signal.shape[0] - n >= hop:
# SCNR in frequency domain
stft.analysis(
noisy_signal[
n : (n + hop),
]
)
gain_filt = scnr.compute_gain_filter(stft.X)
# back to time domain
processed_audio[
n : n + hop,
] = stft.synthesis(gain_filt * stft.X)
# update step
n += hop
return processed_audio
def with_half_overlap_with_filter(D,
num_frames=1,
fixed_memory=False,
streaming=True):
"""
D - number of channels
num_frames - how many frames to process, None will process one frame at
a time
fixed_memory - whether to enforce checks for size (real-time consideration)
streaming - whether or not to stitch between frames
"""
if D == 1:
x_local = x[:, 0]
y_local = y[:, 0]
h_local = h[:, 0]
else:
x_local = x[:, :D]
y_local = y[:, :D]
h_local = h[:, :D]
# parameters
block_size = 512 - h_len + 1 # make sure the FFT size is a power of 2
hop = block_size // 2 # half overlap
window = pra.hann(block_size) # the analysis window
if not streaming:
num_samples = (num_frames - 1) * hop + block_size
x_local = x_local[:num_samples, ]
# Create the STFT object
if fixed_memory:
stft = STFT(block_size,
hop=hop,
channels=D,
transform=transform,
num_frames=num_frames,
analysis_window=window,
streaming=streaming)
else:
stft = STFT(block_size,
hop=hop,
channels=D,
transform=transform,
analysis_window=window,
streaming=streaming)
# setup the filter
stft.set_filter(h_local, zb=h_len - 1)
# collect the processed blocks
processed_x = np.zeros(x_local.shape)
if not streaming:
stft.analysis(x_local)
stft.process()
processed_x = stft.synthesis()
n = processed_x.shape[0]
error = np.max(
np.abs(y_local[block_size:n - block_size, ] -
processed_x[block_size:n - block_size, ]))
else:
n = 0
hop_frames = hop * num_frames
# process the signals while full blocks are available
while x_local.shape[0] - n > hop_frames:
stft.analysis(x_local[n:n + hop_frames, ])
stft.process() # apply the filter
processed_x[n:n + hop_frames, ] = stft.synthesis()
n += hop_frames
error = np.max(np.abs(y_local[:n - hop, ] - processed_x[hop:n, ]))
# if D==1:
# import matplotlib.pyplot as plt
# plt.figure()
# plt.plot(y_local)
# plt.plot(processed_x)
# plt.show()
return error
def with_arbitrary_overlap_synthesis_window(D,
num_frames=1,
fixed_memory=False,
streaming=True,
overlap=0.5):
"""
D - number of channels
num_frames - how many frames to process, None will process one frame at
a time
fixed_memory - whether to enforce checks for size (real-time consideration)
streaming - whether or not to stitch between frames
"""
if D == 1:
x_local = x[:, 0]
else:
x_local = x[:, :D]
# parameters
block_size = 512 # make sure the FFT size is a power of 2
hop = int((1 - overlap) * block_size) # quarter overlap
if not streaming:
num_samples = (num_frames - 1) * hop + block_size
x_local = x_local[:num_samples, ]
analysis_window = pra.hann(block_size)
synthesis_window = pra.transform.compute_synthesis_window(
analysis_window, hop)
# Create the STFT object
if fixed_memory:
stft = STFT(block_size,
hop=hop,
channels=D,
transform=transform,
num_frames=num_frames,
analysis_window=analysis_window,
synthesis_window=synthesis_window,
streaming=streaming)
else:
stft = STFT(block_size,
hop=hop,
channels=D,
analysis_window=analysis_window,
synthesis_window=synthesis_window,
transform=transform,
streaming=streaming)
# collect the processed blocks
processed_x = np.zeros(x_local.shape)
if streaming:
n = 0
hop_frames = hop * num_frames
# process the signals while full blocks are available
while x_local.shape[0] - n > hop_frames:
stft.analysis(x_local[n:n + hop_frames, ])
processed_x[n:n + hop_frames, ] = stft.synthesis()
n += hop_frames
error = np.max(
np.abs(x_local[:n - block_size + hop, ] -
processed_x[block_size - hop:n, ]))
if 20 * np.log10(error) > -10:
import matplotlib.pyplot as plt
if x_local.ndim == 1:
plt.plot(x_local[:n - block_size + hop])
plt.plot(processed_x[block_size - hop:n])
else:
plt.plot(x_local[:n - block_size + hop, 0])
plt.plot(processed_x[block_size - hop:n, 0])
plt.show()
else:
stft.analysis(x_local)
processed_x = stft.synthesis()
n = processed_x.shape[0]
L = block_size - hop
error = np.max(np.abs(x_local[L:-L, ] - processed_x[L:, ]))
if 20 * np.log10(error) > -10:
import matplotlib.pyplot as plt
if x_local.ndim == 1:
plt.plot(x_local[L:-L])
plt.plot(processed_x[L:])
else:
plt.plot(x_local[L:-L, 0])
plt.plot(processed_x[L:, 0])
plt.show()
return error
def apply_iterative_wiener(noisy_signal,
frame_len=512,
lpc_order=20,
iterations=2,
alpha=0.8,
thresh=0.01):
"""
One-shot function to apply iterative Wiener filtering for denoising.
Parameters
----------
noisy_signal : numpy array
Real signal in time domain.
frame_len : int
Frame length in samples. 50% overlap is used with hanning window.
lpc_order : int
Number of LPC coefficients to compute
iterations : int
How many iterations to perform in updating the Wiener filter for each
signal frame.
alpha : int
Smoothing factor within [0,1] for updating noise level. Closer to `1`
gives more weight to the previous noise level, while closer to `0`
gives more weight to the current frame's level. Closer to `0` can track
more rapid changes in the noise level. However, if a speech frame is
incorrectly identified as noise, you can end up removing desired
speech.
thresh : float
Threshold to distinguish between (signal+noise) and (noise) frames. A
high value will classify more frames as noise but might remove desired
signal!
Returns
-------
numpy array
Enhanced/denoised signal.
"""
from pyroomacoustics import hann
from pyroomacoustics.transform import STFT
hop = frame_len // 2
window = hann(frame_len, flag='asymmetric', length='full')
stft = STFT(frame_len, hop=hop, analysis_window=window, streaming=True)
scnr = IterativeWiener(frame_len, lpc_order, iterations, alpha, thresh)
processed_audio = np.zeros(noisy_signal.shape)
n = 0
while noisy_signal.shape[0] - n >= hop:
# SCNR in frequency domain
stft.analysis(noisy_signal[n:(n + hop), ])
X = scnr.compute_filtered_output(current_frame=stft.fft_in_buffer,
frame_dft=stft.X)
# back to time domain
processed_audio[n:n + hop, ] = stft.synthesis(X)
# update step
n += hop
return processed_audio
One frame at a time
"""
print("Averaging computation time over %d cases of %d channels of %d samples (%0.1f s at %0.1f kHz)."
% (num_times,num_mic,len(signals),(len(signals)/fs),fs/1000) )
print()
print("----- SINGLE FRAME AT A TIME -----")
print("With STFT object (not fixed) : ", end="")
stft = STFT(block_size, hop=hop, channels=num_mic,
streaming=True, analysis_window=win)
start = time.time()
for k in range(num_times):
x_r = np.zeros(signals.shape)
n = 0
while signals.shape[0] - n > hop:
stft.analysis(signals[n:n+hop,])
x_r[n:n+hop,] = stft.synthesis()
n += hop
avg_time = (time.time()-start)/num_times
print("%0.3f sec" % avg_time)
err_dB = 20*np.log10(np.max(np.abs(signals[:n-hop,] - x_r[hop:n,])))
print("Error [dB] : %0.3f" % err_dB)
print("With STFT object (fixed) : ", end="")
stft = STFT(block_size, hop=hop, channels=num_mic, num_frames=1,
streaming=True, analysis_window=win)
start = time.time()
for k in range(num_times):
x_r = np.zeros(signals.shape)