def logspec(signal, samplerate, conf):
'''
Compute log magnitude spectrogram features from an audio signal.
Args:
signal: the audio signal from which to compute features. Should be an
N*1 array
samplerate: the samplerate of the signal we are working with.
conf: feature configuration
Returns:
A numpy array of size (NUMFRAMES by numfreq) containing features. Each
row holds 1 feature vector, a numpy vector containing the log magnitude
spectrum of the corresponding frame
'''
signal = sigproc.preemphasis(signal, float(conf['preemph']))
winfunc = _get_winfunc(conf['winfunc'])
frames = sigproc.framesig(signal,
float(conf['winlen']) * samplerate,
float(conf['winstep']) * samplerate, winfunc)
logspec = sigproc.logmagspec(frames, int(conf['nfft']))
return logspec
def fbank(signal,
samplerate=16000,
winlen=0.025,
winstep=0.01,
nfilt=26,
nfft=512,
lowfreq=0,
highfreq=None,
preemph=0.97,
winfunc=lambda x: np.ones((x, ))):
highfreq = highfreq or samplerate / 2
signal = sigproc.preemphasis(signal, preemph)
frames = sigproc.framesig(signal, winlen * samplerate,
winstep * samplerate, winfunc)
pspec = sigproc.powspec(frames, nfft)
energy = np.sum(pspec, 1) # this stores the total energy in each frame
energy = np.where(
energy == 0,
np.finfo(float).eps,
energy) # if energy is zero, we get problems with log
fb = get_filterbanks(nfilt, nfft, samplerate, lowfreq, highfreq)
feat = np.dot(pspec, fb.T) # compute the filterbank energies
feat = np.where(feat == 0,
np.finfo(float).eps,
feat) # if feat is zero, we get problems with log
return feat, energy
def fbank(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97):
"""Compute Mel-filterbank energy features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between seccessive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 20.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:returns: 2 values. The first is a numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector. The
second return value is the energy in each frame (total energy, unwindowed)
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate)
pspec = sigproc.powspec(frames,nfft)
energy = numpy.sum(pspec,1) # this stores the total energy in each frame
fb = get_filterbanks(nfilt,nfft,samplerate)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
return feat,energy
def angspec(signal, samplerate, conf):
"""
Compute angular spectrogram features from an audio signal.
Args:
signal: the audio signal from which to compute features. Should be an
N*1 array
samplerate: the samplerate of the signal we are working with.
conf: feature configuration
Returns:
A numpy array of size (NUMFRAMES by numfreq) containing features. Each
row holds 1 feature vector, a numpy vector containing the angular
spectrum of the corresponding frame
"""
raise BaseException('Not yet implemented')
signal = sigproc.preemphasis(signal, float(conf['preemph']))
winfunc = _get_winfunc(conf['winfunc'])
frames = sigproc.framesig(signal,
float(conf['winlen']) * samplerate,
float(conf['winstep']) * samplerate, winfunc)
angspec = sigproc.angspec(frames, int(conf['nfft']))
return angspec
def RT_CNN():
print("Loading model weights from [{}]....".format(c.WEIGHTS_FILE))
model = vggvox_model() # Creates a VGGVox model
model.load_weights(
c.WEIGHTS_FILE) # Load the weights of the trained models
model.summary() # Print a summary of the loaded model
print("Loading embeddings from enroll")
toLoad = load("data/model/RTSP_CNN.out")
enroll_embs = []
speakers = []
for spk, embs in toLoad.items():
for e in embs:
enroll_embs.append(e)
speakers.append(spk)
print(spk)
count = 0
buffer = AudioBuffer()
start_time = time.time()
while count < 3:
count += 1
buffer.record(chunk_size=c.SAMPLE_RATE)
data = buffer.get_data()
data = np.frombuffer(data, 'int16')
buckets = build_buckets(c.MAX_SEC, c.BUCKET_STEP, c.FRAME_STEP)
data *= 2**15
while (len(data) / (c.FRAME_STEP * c.SAMPLE_RATE) < 101):
data = np.append(data, 0)
# get FFT spectrum
data = remove_dc_and_dither(data, c.SAMPLE_RATE)
data = sigproc.preemphasis(data, coeff=c.PREEMPHASIS_ALPHA)
frames = sigproc.framesig(data,
frame_len=c.FRAME_LEN * c.SAMPLE_RATE,
frame_step=c.FRAME_STEP * c.SAMPLE_RATE,
winfunc=np.hamming)
fft = abs(np.fft.fft(frames, n=c.NUM_FFT))
fft_norm = normalize_frames(fft.T)
# truncate to max bucket sizes
rsize = max(k for k in buckets if k <= len(fft_norm.T))
rstart = int((len(fft_norm.T) - rsize) / 2)
x = fft_norm[:, rstart:rstart + rsize]
test_embs = np.squeeze(model.predict(x.reshape(1, *x.shape, 1)))
distances = []
for embs in enroll_embs:
distances.append(euclidean(test_embs, embs))
print(len(speakers))
idx = np.argmin(distances)
print(speakers[idx])
print("Ok, ", time.time() - start_time - 3, " seconds")
def get_fft_spectrum(filename, buckets):
signal = load_wav(filename, c.SAMPLE_RATE)
signal *= 2**15
# get FFT spectrum (applying hamming)
signal = remove_dc_and_dither(signal, c.SAMPLE_RATE)
signal = sigproc.preemphasis(signal, coeff=c.PREEMPHASIS_ALPHA)
frames = sigproc.framesig(signal,
frame_len=c.FRAME_LEN * c.SAMPLE_RATE,
frame_step=c.FRAME_STEP * c.SAMPLE_RATE,
winfunc=np.hamming)
fft = abs(np.fft.fft(frames, n=c.NUM_FFT))
fft_norm = normalize_frames(
fft.T
) #TODO may remove variance normalization to see other poorer results
# truncate to max bucket sizes
rsize = max(k for k in buckets if k <= fft_norm.shape[1])
rstart = int((fft_norm.shape[1] - rsize) / 2)
out = fft_norm[:, rstart:rstart + rsize]
# print("fft_spectrum shape{}".format(out.shape))
# if(out.shape[1] == c.MAX_SEC*100):
# save_fft_spectrum(filename,out)
return out
def ssc(signal,
samplerate=16000,
winlen=0.025,
winstep=0.01,
nfilt=26,
nfft=512,
lowfreq=0,
highfreq=None,
preemph=0.97,
winfunc=lambda x: np.ones((x, ))):
highfreq = highfreq or samplerate / 2
signal = sigproc.preemphasis(signal, preemph)
frames = sigproc.framesig(signal, winlen * samplerate,
winstep * samplerate, winfunc)
pspec = sigproc.powspec(frames, nfft)
pspec = np.where(pspec == 0,
np.finfo(float).eps,
pspec) # if things are all zeros we get problems
fb = get_filterbanks(nfilt, nfft, samplerate, lowfreq, highfreq)
feat = np.dot(pspec, fb.T) # compute the filterbank energies
R = np.tile(np.linspace(1, samplerate / 2, np.size(pspec, 1)),
(np.size(pspec, 0), 1))
return np.dot(pspec * R, fb.T) / feat
def ssc(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97,
winfunc=lambda x:np.ones((x,))):
"""Compute Spectral Subband Centroid features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the sample rate of the signal we are working with, in Hz.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:param winfunc: the analysis window to apply to each frame. By default no window is applied. You can use numpy window functions here e.g. winfunc=numpy.hamming
:returns: A numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector.
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate, winfunc)
pspec = sigproc.powspec(frames,nfft)
pspec = np.where(pspec == 0,np.finfo(float).eps,pspec) # if things are all zeros we get problems
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = np.dot(pspec,fb.T) # compute the filterbank energies
R = np.tile(np.linspace(1,samplerate/2,np.size(pspec,1)),(np.size(pspec,0),1))
return np.dot(pspec*R,fb.T) / feat
def fbank(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97,
winfunc=lambda x:np.ones((x,))):
"""Compute Mel-filterbank energy features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the sample rate of the signal we are working with, in Hz.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:param winfunc: the analysis window to apply to each frame. By default no window is applied. You can use numpy window functions here e.g. winfunc=numpy.hamming
:returns: 2 values. The first is a numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector. The
second return value is the energy in each frame (total energy, unwindowed)
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate, winfunc)
pspec = sigproc.powspec(frames,nfft)
energy = np.sum(pspec,1) # this stores the total energy in each frame
energy = np.where(energy == 0,np.finfo(float).eps,energy) # if energy is zero, we get problems with log
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = np.dot(pspec,fb.T) # compute the filterbank energies
feat = np.where(feat == 0,np.finfo(float).eps,feat) # if feat is zero, we get problems with log
return feat,energy
def ssc(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97):
"""Compute Spectral Subband Centroid features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between seccessive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 20.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:returns: A numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector.
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate)
pspec = sigproc.powspec(frames,nfft)
fb = get_filterbanks(nfilt,nfft,samplerate)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
R = numpy.tile(numpy.linspace(1,samplerate/2,numpy.size(pspec,1)),(numpy.size(pspec,0),1))
return numpy.dot(pspec*R,fb.T) / feat
def get_fft_spectrum(filename, start, end):
signal = load_wav(filename, c.SAMPLE_RATE)
signal *= 2**15
# get FFT spectrum
signal = remove_dc_and_dither(signal, c.SAMPLE_RATE) # 数字滤波器,去除直流和颤动成分
signal = sigproc.preemphasis(signal,
coeff=c.PREEMPHASIS_ALPHA) # 对输入信号进行预加重
frames = sigproc.framesig(signal,
frame_len=c.FRAME_LEN * c.SAMPLE_RATE,
frame_step=c.FRAME_STEP * c.SAMPLE_RATE,
winfunc=np.hamming) # 将信号框成重叠帧
# print("===================")
# print(frames.shape)
# print("===================")
# exit(0)
spem = sigproc.logpowspec(frames, c.NUM_FFT) # 计算语谱图
# print("===================")
# print(spem)
# print("===================")
# print(spem.shape)
# print("===================")
# exit(0)
spem_norm = normalize_frames(spem.T) # 减去均值，除以标准差
length = spem_norm.shape[1]
reserve_length = length - (length % 100)
# out = fft_norm[:,0:reserve_length] # test
out = spem_norm[:, start:end] # train
return out
def get_fft_spectrum(filename, buckets):
# load the signal with librosa
signal = load_wav(filename, c.SAMPLE_RATE)
# multiply the signal to get the power of 2?
signal *= 2**15
# get FFT spectrum
# not sure what functions below do
signal = remove_dc_and_dither(signal, c.SAMPLE_RATE)
signal = sigproc.preemphasis(signal, coeff=c.PREEMPHASIS_ALPHA)
# build frames to pass in fft
frames = sigproc.framesig(signal,
frame_len=c.FRAME_LEN * c.SAMPLE_RATE,
frame_step=c.FRAME_STEP * c.SAMPLE_RATE,
winfunc=np.hamming)
# get fft spetctrum
fft = abs(np.fft.fft(frames, n=c.NUM_FFT))
# normalize each frame by mean and std
fft_norm = normalize_frames(fft.T)
# truncate to max bucket sizes
# ????
rsize = max(k for k in buckets if k <= fft_norm.shape[1])
rstart = int((fft_norm.shape[1] - rsize) / 2)
out = fft_norm[:, rstart:rstart + rsize]
return out
def fbank(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97):
"""Compute Mel-filterbank energy features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:returns: 2 values. The first is a numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector. The
second return value is the energy in each frame (total energy, unwindowed)
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate)
pspec = sigproc.powspec(frames,nfft)
energy = numpy.sum(pspec,1) # this stores the total energy in each frame
energy = numpy.where(energy == 0,numpy.finfo(float).eps,energy) # if energy is zero, we get problems with log
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
feat = numpy.where(feat == 0,numpy.finfo(float).eps,feat) # if feat is zero, we get problems with log
return feat,energy
def get_fft_spectrum(filename, buckets):
signal = load_wav(filename, c.SAMPLE_RATE)
signal *= 2**15
# print(filename)
print(buckets)
while (len(signal) / (c.FRAME_STEP * c.SAMPLE_RATE) < 101):
signal = np.append(signal, 0)
# get FFT spectrum
signal = remove_dc_and_dither(signal, c.SAMPLE_RATE)
signal = sigproc.preemphasis(signal, coeff=c.PREEMPHASIS_ALPHA)
frames = sigproc.framesig(signal,
frame_len=c.FRAME_LEN * c.SAMPLE_RATE,
frame_step=c.FRAME_STEP * c.SAMPLE_RATE,
winfunc=np.hamming)
fft = abs(np.fft.fft(frames, n=c.NUM_FFT))
# print(len(fft))
fft_norm = normalize_frames(fft.T)
# print(len(fft_norm.T))
# truncate to max bucket sizes
rsize = max(k for k in buckets if k <= len(fft_norm.T))
# print(rsize)
rstart = int((len(fft_norm.T) - rsize) / 2)
# print(rstart)
out = fft_norm[:, rstart:rstart + rsize]
# print(len(out))
return out
def ssc(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97):
"""Compute Spectral Subband Centroid features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:returns: A numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector.
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate)
pspec = sigproc.powspec(frames,nfft)
pspec = numpy.where(pspec == 0,numpy.finfo(float).eps,pspec) # if things are all zeros we get problems
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
R = numpy.tile(numpy.linspace(1,samplerate/2,numpy.size(pspec,1)),(numpy.size(pspec,0),1))
return numpy.dot(pspec*R,fb.T) / feat
def read_and_process_audio(filename, buckets):
signal = read_audio(filename, c.SAMPLE_RATE)
# # Filter out non-speech frequencies
# lowcut, highcut = c.FILTER_RANGE
# signal = butter_bandpass_filter(signal, lowcut, highcut, c.SAMPLE_RATE, 1)
# # Normalize signal
# signal = normalize(signal)
signal *= 2**15
# Process signal to get FFT spectrum
signal = rm_dc_n_dither(signal, c.SAMPLE_RATE)
signal = sigproc.preemphasis(signal, coeff=c.PREEMPHASIS_ALPHA)
frames = sigproc.framesig(signal,
frame_len=c.FRAME_LEN * c.SAMPLE_RATE,
frame_step=c.FRAME_STEP * c.SAMPLE_RATE,
winfunc=np.hamming)
fft = abs(np.fft.fft(frames, n=c.NUM_FFT))
fft_norm = normalize_frames(fft.T)
# Truncate to middle MAX_SEC seconds
rsize = max(k for k in buckets if k <= fft_norm.shape[1])
rstart = int((fft_norm.shape[1] - rsize) / 2)
out = fft_norm[:, rstart:rstart + rsize]
return out
def fbank(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97):
"""Compute Mel-filterbank energy features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:returns: 2 values. The first is a numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector. The
second return value is the energy in each frame (total energy, unwindowed)
"""
highfreq= highfreq or samplerate/2
print "preemph %s"%(preemph)
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate)
matchframes(frames[0], frames[1])
pspec = sigproc.powspec(frames,nfft)
energy = pylab.sum(pspec,1) # this stores the total energy in each frame
energy = pylab.where(energy == 0, pylab.finfo(float).eps, energy) # if energy is zero, we get problems with log
fb = get_filterbanks(nfilt, nfft, samplerate, lowfreq, highfreq)
print "len(fb) %s"%(len(fb))
colour = "k-"
for i in range(len(fb)):
if colour == "k-":
colour = "r-"
else:
colour = "k-"
startedplot = False
midpoint = 0
for j in range(len(fb[i])):
if fb[i][j] > 0:
if startedplot == False:
startedplot = j
if j > 0:
pylab.plot([j-1, j], [fb[i][j-1], fb[i][j]], colour)
if fb[i][j] == 1.0:
midpoint = j
else:
if not startedplot == False:
pylab.plot([j-1, j], [fb[i][j-1], 0], colour)
try:
print "slope to midpoint %.3f, slope from midpoint %.3f"%(1.0/float(midpoint-startedplot), 1.0/float(midpoint-j+1))
except:
pass
break
pylab.show()
feat = pylab.dot(pspec, fb.T) # compute the filterbank energies
feat = pylab.where(feat == 0, pylab.finfo(float).eps, feat) # if feat is zero, we get problems with log
return feat, energy
def vad(sig, rate, winlen, winstep):
'''do voice activity detection
args:
sig: the input signal as a numpy array
rate: the sampling rate
winlen: the window length
winstep: the window step
Returns:
a numpy array of indices containing speech frames
'''
#apply preemphasis
sig = sigproc.preemphasis(sig, 0.97)
#do windowing windowing
frames = sigproc.framesig(sig, winlen * rate, winstep * rate)
#compute the squared frames and center them around zero mean
sqframes = np.square(frames)
sqframes = sqframes - sqframes.mean(1, keepdims=True)
#compute the cross correlation between the frames and their square
corr = np.array(map(partial(np.correlate, mode='same'), frames, sqframes))
#compute the mel power spectrum of the correlated signal
corrfft = np.fft.rfft(corr, 512)
fb = base.get_filterbanks(26, 512, rate, 0, rate / 2)
E = np.absolute(np.square(corrfft).dot(fb.T))
#do noise sniffing at the front and the back and select the lowest energy
Efront = E[:20, :].mean(0)
Eback = E[-20:, :].mean(0)
if Efront.sum() < Eback.sum():
Enoise = Efront
else:
Enoise = Eback
#at every interval compute the mean ratio between the maximal energy in that
#interval and the noise energy
width = 12
#apply max pooling to the energy
Emax = maximum_filter(E, size=[width, 1], mode='constant')
#compute the ratio between the smoothed energy and the noise energy
ratio = np.log((Emax / Enoise).mean(axis=1))
ratio = ratio / np.max(ratio)
speechframes = np.where(ratio > 0.2)[0]
return speechframes
def ssc(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97):
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate)
pspec = sigproc.powspec(frames,nfft)
pspec = numpy.where(pspec == 0,numpy.finfo(float).eps,pspec) # if things are all zeros we get problems
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
R = numpy.tile(numpy.linspace(1,samplerate/2,numpy.size(pspec,1)),(numpy.size(pspec,0),1))
return numpy.dot(pspec*R,fb.T) / feat
def fbank(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97):
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate)
pspec = sigproc.powspec(frames,nfft)
energy = numpy.sum(pspec,1) # this stores the total energy in each frame
energy = numpy.where(energy == 0,numpy.finfo(float).eps,energy) # if energy is zero, we get problems with log
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
feat = numpy.where(feat == 0,numpy.finfo(float).eps,feat) # if feat is zero, we get problems with log
return feat,energy
def fbank(signal,
samplerate=16000,
winlen=0.025,
winstep=0.01,
nfilt=26,
nfft=512,
lowfreq=0,
highfreq=None,
preemph=0.95):
"""
Compute Mel-filterbank energy features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.95.
:returns: 2 values. The first is a numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector. The
second return value is the energy in each frame (total energy, unwindowed)
"""
highfreq = highfreq or samplerate / 2
signal = sigproc.preemphasis(signal, preemph)
# print type(signal[0])
frames = sigproc.framesig(signal,
winlen * samplerate,
winstep * samplerate,
winfunc=hamming_window)
powspec = sigproc.powspec(frames, nfft)
# numpy.savetxt("result.txt", powspec, delimiter=",")
energy = numpy.sum(powspec,
1) # this stores the total energy in each frame
energy = numpy.where(
energy == 0,
numpy.finfo(float).eps, energy
) # if energy is zero, we get problems with log, use numpy.finfo(float).eps to replace 0
filterbanks = get_filterbanks(nfilt, nfft, samplerate, lowfreq, highfreq)
# print powspec.shape, filterbanks.shape
feat = numpy.dot(powspec, filterbanks.T) # compute the filterbank energies
feat = numpy.where(feat == 0,
numpy.finfo(float).eps,
feat) # if feat is zero, we get problems with logs
# print feat.shape
return feat, energy
def get_fft_spectrum(signal):
# padding zero
n_sample = signal.shape[0]
singal_len = int(c.DURA*c.SR)
if n_sample < singal_len:
signal = np.hstack((signal,np.zeros(singal_len-n_sample)))
else:
signal = signal[(n_sample-singal_len)//2:(n_sample+singal_len)//2]
signal = np.array(signal)
signal *= 2**15
signal = remove_dc_and_dither(signal,c.SR)
signal = sigproc.preemphasis(signal,coeff=c.PREEMPHASIS_ALPHA)
frames = sigproc.framesig(signal,frame_len=c.FRAME_LEN*c.SR,frame_step=c.FRAME_STEP*c.SR,winfunc=np.hamming)
fft = abs(np.fft.fft(frames,n= c.N_FFT))
fft_norm = normalization_frames(fft.T)
return fft_norm
def fbankVTLP(signal,
samplerate=16000,
winlen=0.025,
winstep=0.01,
nfilt=26,
nfft=512,
lowfreq=0,
highfreq=None,
preemph=0.97,
appendEnergy=False,
alpha=1.0):
"""Compute Mel-filterbank energy features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:returns: 2 values. The first is a numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector. The
second return value is the energy in each frame (total energy, unwindowed)
"""
highfreq = highfreq or samplerate / 2
signal = sigproc.preemphasis(signal, preemph)
frames = sigproc.framesig(signal, winlen * samplerate,
winstep * samplerate)
pspec = sigproc.powspec(frames, nfft)
energy = numpy.sum(pspec, 1) # this stores the total energy in each frame
energy = numpy.where(energy == 0,
numpy.finfo(float).eps,
energy) # if energy is zero, we get problems with log
fb = get_filterbanksVTLP(nfilt, nfft, samplerate, lowfreq, highfreq, alpha)
feat = numpy.dot(pspec, fb.T) # compute the filterbank energies
feat = numpy.where(feat == 0,
numpy.finfo(float).eps,
feat) # if feat is zero, we get problems with log
if appendEnergy:
feat = numpy.c_[feat, numpy.log(
energy
)] # replace first cepstral coefficient with log of frame energy
return feat, energy
def get_fft_spectrum(signal):
signal *= 2**15
# get FFT spectrum
signal = remove_dc_and_dither(signal, c.SAMPLE_RATE)
signal = sigproc.preemphasis(signal, coeff=c.PREEMPHASIS_ALPHA)
frames = sigproc.framesig(signal,
frame_len=c.FRAME_LEN * c.SAMPLE_RATE,
frame_step=c.FRAME_STEP * c.SAMPLE_RATE,
winfunc=np.hamming)
fft = abs(np.fft.fft(frames, n=c.NUM_FFT))
fft_norm = normalize_frames(fft.T)
rsize = 500
rstart = int((fft_norm.shape[1] - rsize) / 2)
out = fft_norm[:, rstart:rstart + rsize]
return out
def fbank(signal, samplerate, conf):
'''
Compute fbank features from an audio signal.
从一个声音信号中计算fbank特征向量
Args:
参数：
signal: the audio signal from which to compute features. Should be an
N*1 array
要计算特征的声音信号，一个N*1维的数组
samplerate: the samplerate of the signal we are working with.
要处理信号的采样率
conf: feature configuration
特征的配置
Returns:
返回值：
A numpy array of size (NUMFRAMES by nfilt) containing features, a numpy
vector containing the signal energy
返回一个包含特征向量的numpy数组，一个包含信号能量的numpy向量
'''
highfreq = int(conf['highfreq'])
if highfreq < 0:
highfreq = samplerate/2
signal = sigproc.preemphasis(signal, float(conf['preemph']))
frames = sigproc.framesig(signal, float(conf['winlen'])*samplerate,
float(conf['winstep'])*samplerate)
pspec = sigproc.powspec(frames, int(conf['nfft']))
# this stores the total energy in each frame
energy = numpy.sum(pspec, 1)
# if energy is zero, we get problems with log
energy = numpy.where(energy == 0, numpy.finfo(float).eps, energy)
filterbank = get_filterbanks(int(conf['nfilt']), int(conf['nfft']),
samplerate, int(conf['lowfreq']), highfreq)
# compute the filterbank energies
feat = numpy.dot(pspec, filterbank.T)
# if feat is zero, we get problems with log
feat = numpy.where(feat == 0, numpy.finfo(float).eps, feat)
return feat, energy
def get_fft_spectrum(filename, buckets): signal = load_wav(filename,c.SAMPLE_RATE) signal *= 2**15 # get FFT spectrum signal = remove_dc_and_dither(signal, c.SAMPLE_RATE) signal = sigproc.preemphasis(signal, coeff=c.PREEMPHASIS_ALPHA) frames = sigproc.framesig(signal, frame_len=c.FRAME_LEN*c.SAMPLE_RATE, frame_step=c.FRAME_STEP*c.SAMPLE_RATE, winfunc=np.hamming) fft = abs(np.fft.fft(frames,n=c.NUM_FFT)) fft_norm = normalize_frames(fft.T) # truncate to max bucket sizes rsize = max(k for k in buckets if k <= fft_norm.shape[1]) rstart = int((fft_norm.shape[1]-rsize)/2) out = fft_norm[:,rstart:rstart+rsize] # print(out.shape) # exit(0) return out
def fbank(signal, samplerate, conf):
"""
Compute fbank features from an audio signal.
Args:
signal: the audio signal from which to compute features. Should be an
N*1 array
samplerate: the samplerate of the signal we are working with.
conf: feature configuration
Returns:
A numpy array of size (NUMFRAMES by nfilt) containing features, a numpy
vector containing the signal energy
"""
raise BaseException('Not yet implemented')
highfreq = int(conf['highfreq'])
if highfreq < 0:
highfreq = samplerate / 2
signal = sigproc.preemphasis(signal, float(conf['preemph']))
frames = sigproc.framesig(signal,
float(conf['winlen']) * samplerate,
float(conf['winstep']) * samplerate)
pspec = sigproc.powspec(frames, int(conf['nfft']))
# this stores the total energy in each frame
energy = numpy.sum(pspec, 1)
# if energy is zero, we get problems with log
energy = numpy.where(energy == 0, numpy.finfo(float).eps, energy)
filterbank = get_filterbanks(int(conf['nfilt']), int(conf['nfft']),
samplerate, int(conf['lowfreq']), highfreq)
# compute the filterbank energies
feat = numpy.dot(pspec, filterbank.T)
# if feat is zero, we get problems with log
feat = numpy.where(feat == 0, numpy.finfo(float).eps, feat)
return feat, energy
def angspec_scipy(signal, samplerate, conf):
"""
Compute angular spectrogram features from an audio signal.
Args:
signal: the audio signal from which to compute features. Should be an
N*1 array
samplerate: the samplerate of the signal we are working with.
conf: feature configuration
Returns:
A numpy array of size (NUMFRAMES by numfreq) containing features. Each
row holds 1 feature vector, a numpy vector containing the angular
spectrum of the corresponding frame
"""
signal = sigproc.preemphasis(signal, float(conf['preemph']))
spec = spec_scipy(signal, samplerate, conf)
angspec = np.angle(spec)
return angspec
def get_fft_spectrum(filename, buckets):
signal = load_wav(filename, 16000)
signal *= 2**15
# get FFT spectrum
signal = remove_dc_and_dither(signal, 16000)
signal = sigproc.preemphasis(signal, coeff=0.97)
frames = sigproc.framesig(signal,
frame_len=0.025 * 16000,
frame_step=0.01 * 16000,
winfunc=np.hamming)
fft = abs(np.fft.fft(frames, n=512))
fft_norm = normalize_frames(fft.T)
# truncate to max bucket sizes
rsize = max(k for k in buckets if k <= fft_norm.shape[1])
rstart = int((fft_norm.shape[1] - rsize) / 2)
out = fft_norm[:, rstart:rstart + rsize]
return out
def ssc(signal, samplerate, conf):
'''
Compute ssc features from an audio signal.
Args:
signal: the audio signal from which to compute features. Should be an
N*1 array
samplerate: the samplerate of the signal we are working with.
conf: feature configuration
Returns:
A numpy array of size (NUMFRAMES by nfilt) containing features, a numpy
vector containing the signal log-energy
'''
highfreq = int(conf['highfreq'])
if highfreq < 0:
highfreq = samplerate / 2
signal = sigproc.preemphasis(signal, float(conf['preemph']))
frames = sigproc.framesig(signal,
float(conf['winlen']) * samplerate,
float(conf['winstep']) * samplerate)
pspec = sigproc.powspec(frames, int(conf['nfft']))
# this stores the total energy in each frame
energy = numpy.sum(pspec, 1)
# if energy is zero, we get problems with log
energy = numpy.where(energy == 0, numpy.finfo(float).eps, energy)
filterbank = get_filterbanks(int(conf['nfilt']), int(conf['nfft']),
samplerate, int(conf['lowfreq']), highfreq)
# compute the filterbank energies
feat = numpy.dot(pspec, filterbank.T)
tiles = numpy.tile(numpy.linspace(1, samplerate / 2, numpy.size(pspec, 1)),
(numpy.size(pspec, 0), 1))
return numpy.dot(pspec * tiles, filterbank.T) / feat, numpy.log(energy)
def frames(signal, samplerate, conf):
"""
Compute frames from an audio signal.
Args:
signal: the audio signal from which to compute features. Should be an
N*1 array
samplerate: the samplerate of the signal we are working with.
conf: feature configuration
Returns:
A numpy array of size (NUMFRAMES by winlen) containing features. Each
row holds 1 feature vector
"""
signal = sigproc.preemphasis(signal, float(conf['preemph']))
winfunc = _get_winfunc(conf['winfunc'])
frames = sigproc.framesig(signal,
float(conf['winlen']) * samplerate,
float(conf['winstep']) * samplerate, winfunc)
return frames
def do_preemp(self, preamp):
""" apply pre-emphasis """
self.samples = sigproc.preemphasis(self.samples, preamp)
