def eRatio(dirName):
for root, dirs, files in os.walk(str(dirName)):
for file in files:
if file.endswith('.wav'):
wavobj = wavio.read(root + '\\' + file)
sampleRate = wavobj.rate
data = wavobj.data
if data is not 'float':
data = data.astype('float') # data / 32768.0
if np.shape(np.shape(data))[0] > 1:
data = data[:, 0]
post = SupportClasses.postProcess(data, sampleRate, [])
print(file, post.eRatioConfd(seg=None))
def deleteClick2(dirName):
"""
Given the directory of sounds this deletes the annotation segments with rain corrupted.
Check to make sure the segment to delete has no sign of kiwi - use fundamental frq rather than eRatio in 'deleteClick'
"""
for root, dirs, files in os.walk(str(dirName)):
for file in files:
if file.endswith('.data') and file[:-5] in files:
# go through each segment
file = root + '/' + file
with open(file) as f:
segments = json.load(f)
newSegments = copy.deepcopy(segments)
wavobj = wavio.read(file[:-5])
audioData = wavobj.data
if audioData is not 'float':
audioData = audioData / 32768.0
audioData = audioData[:, 0].squeeze()
sampleRate = wavobj.rate
if sampleRate != 16000:
audioData = librosa.core.audio.resample(audioData, sampleRate, 16000)
sampleRate = 16000
# Find T_ERatio based on first 5 secs as it varies accorss the recordings
post = SupportClasses.postProcess(audioData, sampleRate, [])
# T_ERatio = post.eRatioConfd([1, 6, "", ""])
print(file)
if len(segments)>2:
ff = Features.Features(audioData, sampleRate)
mfcc = ff.get_mfcc()
mean = np.mean(mfcc[1, :])
std = np.std(mfcc[1, :])
thr = mean - 2 * std # mfcc1 thr for the file
else:
thr = 0
chg = False
for seg in segments:
if seg[0] == -1:
continue
else:
# read the sound segment and check for wind
secs = seg[1] - seg[0]
wavobj = wavio.read(file[:-5], nseconds=secs, offset=seg[0])
data = wavobj.data
sampleRate =wavobj.rate
if data is not 'float':
data = data / 32768.0
data = data[:, 0].squeeze()
# check for clicks
ff = Features.Features(data, sampleRate)
mfcc = ff.get_mfcc()
mfcc1 = mfcc[1, :] # mfcc1 of the segment
if thr == 0:
ff = Features.Features(audioData, sampleRate)
mfcc = ff.get_mfcc()
mean = np.mean(mfcc[1,:])
std = np.std(mfcc[1,:])
thr = mean - 2 * std # mfcc1 thr for the file
if np.min(mfcc1) < thr:
# # now check eRatio
# eRatio = post.eRatioConfd(seg) # eRatio(file[:-5], seg, thr=T_ERatio)
# eRatioBefore = post.eRatioConfd([seg[0] - 10, seg[0], "", ""])
# eRatioAfter = post.eRatioConfd([seg[1], seg[1] + 10, "", ""])
# if eRatio > eRatioBefore*1.05 or eRatio > eRatioAfter*1.05:
# continue
# now check f. frq.
# down sample to avoid higher frq noise
if sampleRate != 16000:
data = librosa.core.audio.resample(data, sampleRate, 16000)
sampleRate = 16000
# denoise prior to f frq detection
waveletDenoiser = WaveletFunctions.WaveletFunctions(data=data, wavelet=None, maxLevel=12)
data = waveletDenoiser.waveletDenoise(data, thresholdType='soft', wavelet='dmey2', maxLevel=12)
sp = SignalProc.SignalProc([], 0, 512, 256)
sgRaw = sp.spectrogram(data, 512,256, mean_normalise=True, onesided=True, multitaper=False)
segment = Segment.Segment(data, sgRaw, sp, sampleRate,512, 256)
pitch, y, minfreq, W = segment.yin()
ind = np.squeeze(np.where(pitch > minfreq))
pitch = pitch[ind]
ff = np.mean(pitch)
if ff>500 and ff<5000:
continue
else:
print(seg)
newSegments.remove(seg)
chg = True
if chg:
file = open(file, 'w')
json.dump(newSegments, file)
def deleteClick(dirName):
"""
Given the directory of sounds this deletes the annotation segments with wind/rain corrupted files.
Targeting moderate wind and above. Check to make sure the segment to delete has no sign of kiwi
"""
for root, dirs, files in os.walk(str(dirName)):
for file in files:
if file.endswith('.data') and file[:-5] in files:
# go through each segment
file = root + '/' + file
with open(file) as f:
segments = json.load(f)
newSegments = copy.deepcopy(segments)
wavobj = wavio.read(file[:-5])
audioData = wavobj.data
if audioData is not 'float':
audioData = audioData / 32768.0
audioData = audioData[:, 0].squeeze()
sampleRate = wavobj.rate
# Find T_ERatio based on first 5 secs as it varies accorss the recordings
post = SupportClasses.postProcess(audioData, sampleRate, [])
# T_ERatio = post.eRatioConfd([1, 6, "", ""])
print(file)
chg = False
for seg in segments:
if seg[0] == -1:
continue
else:
# read the sound segment and check for wind
secs = seg[1] - seg[0]
wavobj = wavio.read(file[:-5], nseconds=secs, offset=seg[0])
data = wavobj.data
if data is not 'float':
data = data / 32768.0
data = data[:, 0].squeeze()
# check for clicks
ff = Features.Features(data, sampleRate)
mfcc = ff.get_mfcc()
mfcc1 = mfcc[1, :] # mfcc1 of the segment
ff = Features.Features(audioData, sampleRate)
mfcc = ff.get_mfcc()
mean = np.mean(mfcc[1,:])
std = np.std(mfcc[1,:])
thr = mean - 2 * std # mfcc1 thr for the file
if np.min(mfcc1) < thr:
# # # now check eRatio
# eRatio = post.eRatioConfdV2(seg)
# if eRatio > 1.0:
# continue
# just check duration>10 sec
if secs > 10:
continue
else:
print(seg)
newSegments.remove(seg)
chg = True
if chg:
file = open(file, 'w')
json.dump(newSegments, file)
def deleteWindRain(dirName, windTest=True, rainTest=False, Tmean_wind = 1e-8):
"""
Given the directory of sounds this deletes the annotation segments with wind/rain corrupted files.
Targeting moderate wind and above. Check to make sure the segment to delete has no sign of kiwi
Automatic Identification of Rainfall in Acoustic Recordings by Carol Bedoya, Claudia Isaza, Juan M.Daza, and Jose D.Lopez
"""
#Todo: find thrs
Tmean_rain = 1e-8 # Mean threshold
Tsnr_rain = 3.5 # SNR threshold
# Tmean_wind = 1e-9 # Mean threshold
# Tsnr_wind = 0.5 # SNR threshold
cnt = 0
for root, dirs, files in os.walk(str(dirName)):
for file in files:
if file.endswith('.data') and file[:-5] in files:
# go through each segment
file = root + '/' + file
with open(file) as f:
segments = json.load(f)
newSegments=copy.deepcopy(segments)
wavobj = wavio.read(file[:-5])
audioData = wavobj.data
# # ***
# if audioData.dtype is not 'float':
# audioData = audioData.astype('float') # / 32768.0
# if np.shape(np.shape(audioData))[0] > 1:
# audioData = np.squeeze(audioData[:, 0])
# import librosa
# if wavobj.rate != 16000:
# audioData = librosa.core.audio.resample(audioData, wavobj.rate, 16000)
# sampleRate = 16000
# # ****
if audioData is not 'float':
audioData = audioData / 32768.0
audioData = audioData[:, 0].squeeze()
sampleRate = wavobj.rate
# Find T_ERatio based on first 5 secs as it varies accorss the recordings
post = SupportClasses.postProcess(audioData, sampleRate, [])
# T_ERatio = post.eRatioConfd([1, 6, "", ""])
chg = False
for seg in segments:
if seg[0] == -1:
continue
else:
# read the sound segment and check for wind
secs = seg[1]-seg[0]
wavobj = wavio.read(file[:-5], nseconds=secs, offset=seg[0])
data = wavobj.data
# # ***
# if data.dtype is not 'float':
# data = data.astype('float') # / 32768.0
# if np.shape(np.shape(data))[0] > 1:
# data = np.squeeze(data[:, 0])
# if wavobj.rate != 16000:
# data = librosa.core.audio.resample(data, wavobj.rate, 16000)
if data is not 'float':
data = data / 32768.0
data = data[:,0].squeeze()
wind_lower = 2.0 * 100 / sampleRate
wind_upper = 2.0 * 250 / sampleRate
rain_lower = 2.0 * 600 / sampleRate
rain_upper = 2.0 * 1200 / sampleRate
f, p = signal.welch(data, fs=sampleRate, window='hamming', nperseg=512, detrend=False)
if windTest:
limite_inf = int(round(len(p) * wind_lower)) # minimum frequency of the rainfall frequency band 0.00625(in
# normalized frequency); in Hz = 0.00625 * (44100 / 2) = 100 Hz
limite_sup = int(round(len(p) * wind_upper)) # maximum frequency of the rainfall frequency band 0.03125(in
# normalized frequency); in Hz = 0.03125 * (44100 / 2) = 250 Hz
a_wind = p[limite_inf:limite_sup] # section of interest of the power spectral density.Step 2 in Algorithm 2.1
mean_a_wind = np.mean(a_wind) # mean of the PSD in the frequency band of interest.Upper part of the step 3 in Algorithm 2.1
std_a_wind = np.std(a_wind) # standar deviation of the PSD in the frequency band of the interest. Lower part of the step 3 in Algorithm 2.1
# c_wind = mean_a_wind / std_a_wind # signal to noise ratio of the analysed recording. step 3 in Algorithm 2.1
if mean_a_wind > Tmean_wind:
# eRatio = post.eRatioConfd(seg) #eRatio(file[:-5], seg, thr=T_ERatio)
# eRatioBefore = post.eRatioConfd([seg[0]-10, seg[0], "", ""])
# if eRatio > eRatioBefore*1.05: # or eRatio > eRatioAfter: #it was 10 secs Before eratio
# #version2
# eRatio = post.eRatioConfdV2(seg)
# if eRatio > 1.0:
# potentialCall = True
# # now check f. frq.
# # down sample will helkp to avoid higher frq noise
# if sampleRate != 16000:
# data = librosa.core.audio.resample(data, sampleRate, 16000)
# sampleRate = 16000
# # denoise prior to f. frq. detection
# waveletDenoiser = WaveletFunctions.WaveletFunctions(data=data, wavelet=None,
# maxLevel=12)
# data = waveletDenoiser.waveletDenoise(data, thresholdType='soft', wavelet='dmey2',
# maxLevel=12)
# sp = SignalProc.SignalProc([], 0, 512, 256)
# sgRaw = sp.spectrogram(data, 512, 256, mean_normalise=True, onesided=True,
# multitaper=False)
# segment = Segment.Segment(data, sgRaw, sp, sampleRate, 512, 256)
# pitch, y, minfreq, W = segment.yin(minfreq=600)
# ind = np.squeeze(np.where(pitch > minfreq))
# pitch = pitch[ind]
# ff = np.mean(pitch)
# if ff > 500 and ff < 5000:
# potentialCall = True
# else:
# potentialCall = False
# just check duration>10 sec
if secs>10:
potentialCall = True
else:
potentialCall = False
if not potentialCall:
print(file, seg, "--> windy")
newSegments.remove(seg)
chg = True
else:
print(file, seg, "--> not windy")
if rainTest:
limite_inf = int(round(len(p) * rain_lower)) # minimum frequency of the rainfall frequency band 0.0272 (in
# normalized frequency); in Hz=0.0272*(44100/2)=599.8 Hz
limite_sup = int(round(len(p) * rain_upper)) # maximum frequency of the rainfall frequency band 0.0544 (in
# normalized frequency); in Hz=0.0544*(44100/2)=1199.5 Hz
a_rain = p[limite_inf:limite_sup] # section of interest of the power spectral density.Step 2 in Algorithm 2.1
mean_a_rain = np.mean(a_rain) # mean of the PSD in the frequency band of interest.Upper part of the step 3 in Algorithm 2.1
std_a_rain = np.std(a_rain) # standar deviation of the PSD in the frequency band of the interest. Lower part of the step 3 in Algorithm 2.1
c_rain = mean_a_rain / std_a_rain # signal to noise ratio of the analysed recording. step 3 in Algorithm 2.1
if c_rain > Tsnr_rain:
# check if it is not kiwi
eRatio = post.eRatioConfd(seg) # eRatio(file[:-5], seg, thr=T_ERatio)
eRatioBefore = post.eRatioConfd([seg[0] - 10, seg[0], "", ""])
# eRatioAfter = post.eRatioConfd([seg[1], seg[1] + 5, "", ""])
# T_ERatio = (eRatioBefore + eRatioAfter) / 2
if eRatio > eRatioBefore: # or eRatio > eRatioAfter: #it was 10 secs Before eratio
potentialCall = True
else:
potentialCall = False
if not potentialCall:
print(file, seg, "--> windy")
newSegments.remove(seg)
chg = True
else:
# rainy.append(0)
print(file, "--> not rainy")
if chg:
file = open(file, 'w')
json.dump(newSegments, file)
cnt += 1
print(file, cnt)