def correct_and_check(self,path): """ The VTL-Produced sounds often come with short bursts right at the onset of the sound. This little function removes such noises as they interfere with learning. Sometimes we don't have any sound at all. Give that info as output to the main function! passed: "did it pass the test?" """ sound = loadsound(path) #loadsound from brianhears low = 249 #duration of initial silence sound[0:low] = 0 sound.save(path) # Look at the sound at different places: Will they add up to a reasonable number? mean = 0 for here in range(4000,12001,200): mean += abs(sound[here][0]) # Only non-silent sounds are valid. valid = mean!=0. #from matplotlib import pyplot as plt #plt.plot(sound) #plt.show() return valid
def ncc(audio):
# load .wav file through brian.hears.loadsound()
x, _ = librosa.load(audio, sr=16000)
sf.write('tmp.wav', x, 16000)
audio = bh.loadsound('tmp.wav')
t_audio = audio.size / 16000
# set the observing frequences
# human hearing frequency range 20 ~ 20000 Hz
freqs = mel2hz(np.linspace(hz2mel(20), hz2mel(8000), 4))
print freqs
# pass audio to AuditoryPeriphery net
APnet = AudPeri(20, audio, freqs)
# pass AuditoryPeriphery net to Ceptral net
Cepnet = Cep(20, freqs, 4, APnet)
# the whole NCC network is the network returned by Ceptral net
NCC = Cepnet
# dNCC = IntermediateDeriv(20, freqs, Cepnet)
# # pass Ceptral net to Derivative net
# # choose between the two models to simulate the derivative network
# if args.deriv == 'interm' : NCC = IntermediateDeriv(20, freqs, Cepnet)
# # elif args.deriv == 'feedforward' :
# NCC = FeedforwardDeriv(20, freqs, Cepnet)
# the returned net is the NCCs network
with NCC:
# observe the output of NCCs network with a probe
probe = nengo.Probe(NCC.output, synapse=0.01)
def extract_features(fname, bdir, sox, htk_mfc, mfc_extension, stereo_wav,
gammatones, spectrograms, filterbanks):
#def extract_features(fname, bdir):
if fname[-4:] != '.wav':
return
rawfname = bdir+'/'+fname[:-4]+'.rawaudio'
wavfname = bdir+'/'+fname
tempfname = bdir+'/'+fname[:-4]+'_temp.wav'
# temp fname with .wav for sox
mfccfname = bdir+'/'+fname[:-4]+mfc_extension
if sox:
shutil.move(wavfname, tempfname)
call(['sox', tempfname, wavfname])
#call(['sox', '-G', tempfname, '-r 16k', wavfname])
# w/o headers, sox uses extension
shutil.move(tempfname, rawfname)
if htk_mfc:
call(['HCopy', '-C', 'wav_config', wavfname, mfccfname])
srate = 16000
#srate, sound = wavfile.read(wavfname)
sound, srate = readwav(wavfname)
if stereo_wav and len(sound.shape) == 2: # in mono sound is a list
sound = 0.5 * (sound[:, 0] + sound[:, 1])
# for stereo wav, sum both channels
if gammatones:
gammatonefname = bdir+'/'+fname[:-4]+'_gamma.npy'
tmp_snd = loadsound(wavfname)
gamma_cf = erbspace(20*Hz, 20*kHz, N_GAMMATONES_FILTERS)
gamma_fb = Gammatone(tmp_snd, gamma_cf)
with open(gammatonefname, 'w') as o_f:
npsave(o_f, gamma_fb.process())
if spectrograms:
powerspec, _, _, _ = specgram(sound, NFFT=int(srate
* SPECGRAM_WINDOW), Fs=srate, noverlap=int(srate
* SPECGRAM_OVERLAP)) # TODO
specgramfname = bdir+'/'+fname[:-4]+'_specgram.npy'
with open(specgramfname, 'w') as o_f:
npsave(o_f, powerspec.T)
if filterbanks:
# convert to Mel filterbanks
fbanks = Spectral(nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
compression='log',
fs=srate, # sampling rate
lowerf=50, # lower frequency
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
sound /= np.abs(sound).max(axis=0) # TODO put that as option
fbank = fbanks.transform(sound)
fbanksfname = bdir+'/'+fname[:-4]+'_fbanks.npy'
with open(fbanksfname, 'w') as o_f:
npsave(o_f, fbank)
# TODO wavelets scattergrams / scalograms
print "dealt with file", wavfname
def process(folder,
debug=False,
htk_mfc=False,
forcemfcext=False,
stereo_wav=False,
gammatones=False,
spectrograms=False):
""" debug output? HCopy for MFCC? wav are stereo? produce gammatones? """
# first find if we produce normalized MFCC, otherwise note it in the ext
# because we can then normalize on the whole corpus with another py script
mfc_extension = '.mfc_unnorm'
wcfg = open('wav_config', 'r')
for line in wcfg:
if "ENORMALISE" in line:
mfc_extension = '.mfc'
if forcemfcext:
mfc_extension = '.mfc'
print "MFC extension:", mfc_extension
# run through all the folders and files in the path "folder"
# and put a header to the waves, save the originals as .rawaudio
# use HCopy to produce MFCC files according to "wav_config" file
for d, ds, fs in os.walk(folder):
for fname in fs:
if fname[-4:] != '.wav':
continue
rawfname = d+'/'+fname[:-4]+'.rawaudio'
wavfname = d+'/'+fname
tempfname = d+'/'+fname[:-4]+'_temp.wav' # temp fname with .wav for sox
mfccfname = d+'/'+fname[:-4]+mfc_extension
shutil.move(wavfname, tempfname)
call(['sox', tempfname, wavfname]) # w/o headers, sox uses extension
shutil.move(tempfname, rawfname)
if htk_mfc:
call(['HCopy', '-C', 'wav_config', wavfname, mfccfname])
sr = 16000
sr, sound = wavfile.read(wavfname)
if stereo_wav and len(sound.shape) == 2: # in mono sound is a list
sound = sound[:,1] # for stereo wav, arbitrarily take channel 1
if gammatones:
from brian import Hz, kHz
from brian.hears import loadsound, erbspace, Gammatone
gammatonefname = d+'/'+fname[:-4]+'_gamma.npy'
tmp_snd = loadsound(wavfname)
cf = erbspace(20*Hz, 20*kHz, N_GAMMATONES_FILTERS)
fb = Gammatone(tmp_snd, cf)
with open(gammatonefname, 'w') as of:
numpy.save(of, fb.process())
if spectrograms:
from pylab import specgram
Pxx, freqs, bins, im = specgram(sound, NFFT=int(sr * SPECGRAM_WINDOW), Fs=sr, noverlap=int(sr * SPECGRAM_OVERLAP))
specgramfname = d+'/'+fname[:-4]+'_specgram.npy'
with open(specgramfname, 'w') as of:
numpy.save(of, Pxx.T)
print "dealt with file", wavfname
def process(folder,
debug=False,
htk_mfc=False,
forcemfcext=False,
stereo_wav=False,
gammatones=False,
spectrograms=False,
filterbanks=False,
sox=True):
""" applies to all *.wav in folder """
# first find if we produce normalized MFCC, otherwise note it in the ext
# because we can then normalize on the whole corpus with another py script
mfc_extension = '.mfc_unnorm'
wcfg = open('wav_config', 'r')
for line in wcfg:
if "ENORMALISE" in line:
mfc_extension = '.mfc'
if forcemfcext:
mfc_extension = '.mfc'
print "MFC extension:", mfc_extension
if gammatones:
try:
from brian import Hz, kHz
from brian.hears import loadsound, erbspace, Gammatone
except ImportError:
print >> sys.stderr, "You need Brian Hears"
print >> sys.stderr, "http://www.briansimulator.org/docs/\
hears.html"
sys.exit(-1)
if spectrograms:
try:
from pylab import specgram
except ImportError:
print >> sys.stderr, "You need Pylab"
sys.exit(-1)
fbanks = None
if filterbanks:
try:
sys.path.append('../spectral')
from spectral import Spectral
except ImportError:
print >> sys.stderr, "You need spectral (in the parent folder)"
print >> sys.stderr, "https://github.com/mwv/spectral"
sys.exit(-1)
# run through all the folders and files in the path "folder"
# and put a header to the waves, save the originals as .rawaudio
# use HCopy to produce MFCC files according to "wav_config" file
for bdir, _, files in os.walk(folder):
for fname in files:
if fname[-4:] != '.wav':
continue
rawfname = bdir + '/' + fname[:-4] + '.rawaudio'
wavfname = bdir + '/' + fname
tempfname = bdir + '/' + fname[:-4] + '_temp.wav'
# temp fname with .wav for sox
mfccfname = bdir + '/' + fname[:-4] + mfc_extension
if sox:
shutil.move(wavfname, tempfname)
call(['sox', tempfname, wavfname])
# w/o headers, sox uses extension
shutil.move(tempfname, rawfname)
if htk_mfc:
call(['HCopy', '-C', 'wav_config', wavfname, mfccfname])
srate = 16000
srate, sound = wavfile.read(wavfname)
if stereo_wav and len(sound.shape) == 2: # in mono sound is a list
sound = sound[:, 0] + sound[:, 1]
# for stereo wav, sum both channels
if gammatones:
gammatonefname = bdir + '/' + fname[:-4] + '_gamma.npy'
tmp_snd = loadsound(wavfname)
gamma_cf = erbspace(20 * Hz, 20 * kHz, N_GAMMATONES_FILTERS)
gamma_fb = Gammatone(tmp_snd, gamma_cf)
with open(gammatonefname, 'w') as o_f:
npsave(o_f, gamma_fb.process())
if spectrograms:
powerspec, _, _, _ = specgram(
sound,
NFFT=int(srate * SPECGRAM_WINDOW),
Fs=srate,
noverlap=int(srate * SPECGRAM_OVERLAP)) # TODO
specgramfname = bdir + '/' + fname[:-4] + '_specgram.npy'
with open(specgramfname, 'w') as o_f:
npsave(o_f, powerspec.T)
if filterbanks:
# convert to Mel filterbanks
if fbanks == None: # assume parameters are fixed
fbanks = Spectral(
nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
do_dct=False, # we do not want MFCCs
fs=srate, # sampling rate
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
do_deltas=False, # speed
do_deltasdeltas=False # acceleration
)
fbank = fbanks.transform(sound)[0] # first dimension is for
# deltas & deltasdeltas
fbanksfname = bdir + '/' + fname[:-4] + '_fbanks.npy'
with open(fbanksfname, 'w') as o_f:
npsave(o_f, fbank)
# TODO wavelets scattergrams / scalograms
print "dealt with file", wavfname
def process(folder,debug=False,htk_mfcc=False,forcemfcext=False,stereo_wave=False,gammatones=False,spectograms=False,filterbanks=False,sox=True):
mfc_extension = '.mfc_unnorm'
wcfg = open('wav_config','r')
for line in wcfg:
if "ENORMALISE" in line:
mfc_extension = '.mfc'
if forcemfcext:
mfc_extension = '.mfc'
print "MFC Extension is", mfc_extension
if gammatones:
try:
from brian import Hz, kHz
from brian.hears import loadsound, erbspace, Gammatone
except ImportError:
print >> sys.stderr, "You need Brian Hears"
sys.exit(-1)
if spectograms:
try:
from pylab import specgram
except ImportError:
print >> sys.stderr,'You need Pylab'
sys.exit(-1)
fbanks = None
if filterbanks:
try:
sys.path.append('../spectral')
from spectral import Spectral
except ImportError:
print >> sys.stderr, 'you need spectral (in the parent folder)'
for bdir, _ , files in os.walk(folder):
for fname in files:
if fname[-4:] != '.WAV':
continue
rawfname= bdir + '/' + fname[:-4]+'.rawaudio'
wavfname = bdir + '/'+ fname
tempfname = bdir + '/' + fname[:-4] + '_temp.wav'
mfccfname = bdir + '/' + fname[:-4] + '.txt'
if sox:
shutil.move(wavfname, tempfname)
call(['sox',tempfname,wavfname])
shutil.move(tempfname,wavfname)
if htk_mfcc:
call(['HCopy','-C','wav_config',wavfname,mfccfname])
srate = 16000
srate, sound = wavfile.read(wavfname)
if stereo_wave and len(sound.shape == 2):
sound = sound[:,0]+ sound[:,1]
if gammatones:
gammatonefname = bdir + '/' + fname[:-4] + '_gamma.npy'
tmp_snd = loadsound(wavfname)
gamma_cf = erbspace(20*Hz, 20*kHz, n_gmammatones_filters)
gamma_fb = Gammatone(tmp_snd, gamma_cf)
with open(gammatonefname,'w') as o_f:
npsave(o_f, gamma_fb.process())
if spectograms:
powersspec, _,_,_ = specgram(sound, NFFT=int(srate * specgram_window), Fs=srate,noverlap=int(srate*specgram_window))
specgramfname = bdir + '/' + fname[:-4]+'_specgram.npy'
with open(specgramfname,'w') as o_f:
npsave(o_f , powerspec.T)
if filterbanks:
if fbanks ==None:
fbanks = Spectral(nfilt = n_fbanks, alpha=0.97,do_dct=False, fs=srate, frate=fbanks_rate, wlen=fbanks_window,nfft=1024,do_deltas=False,do_deltasdeltas=False)
fbank = fbanks.transform(sound)[0]
fbanksfname = bdir + '/' + fname[:-4]+'_fbanks.npy'
with open(fbanksfname,'w') as o_f:
npsave(o_f, fbank)
print "Dealt with the file ", wavfname
'group_speaker':group_speaker,
'pitch_var':self.speaker_pitch_rel[speaker],
'verbose':True }
paths = { 'input_path':self.learner_path,
'wav_folder':self.output_path+'/'+vowel }
synthesize.main(input_dict,paths)
wavFile = synthesize_wav.main( params, speaker, simulation_name,pitch_var=0,len_var=1.0,verbose=self.verbose,self.rank=self.rank,
different_output_path=self.output_folder)
# wavFile = par_to_wav(params, speaker, simulation_name, verbose=self.verbose, self.rank=self.rank) # call parToWave to generate sound file
if self.verbose:
print 'wav file '+str(wavFile)+' produced'
sound = loadsound(wavFile) # load sound file for brian.hears processing
if self.verbose:
print 'sound loaded'
############### Audio processing
sound = correct_initial(sound) # call correct_initial to remove initial burst
sound_resampled = get_resampled(sound)
# call get_resampled to adapt generated sound to AN model
sound_extended = get_extended(sound_resampled)
# call get_extended to equalize duration of all sounds
sound_extended.save(wavFile) # save current sound as sound file
def extract_from_mlf(mlf, do_gammatones):
x = np.ndarray((0, N_MFCC_COEFFS), dtype='float32')
x_fbank = np.ndarray((0, N_FILTERBANK_COEFFS), dtype='float32')
x_gamma = np.ndarray((0, N_GAMMATONES * 3), dtype='float32')
y = []
y_spkr = []
with open(mlf) as f:
tmp_len_x = 0 # verify sizes
len_x = 0
end = 0
speaker_label = ''
for line in f:
line = line.rstrip('\n')
if len(line) < 1:
continue
if line[0] == '"':
assert tmp_len_x == 0, "the file above this one %s was mismatching x (%d frames) and y (%d frames) lengths by %d" % (
line, len_x, end, tmp_len_x)
speaker_label = line.split('/')[-2]
# load HTK's MFCC
t = htkmfc.open(line.strip('"')[:-3] + 'mfc') # .lab -> .mfc
x = np.append(x, t.getall(), axis=0)
len_x = t.getall().shape[0]
tmp_len_x = len_x
if TALKBOX_FBANKS: # do our own filterbanks TODO
fr, snd = wavfile.read(line.strip('"')[:-3] +
'wav') # .lab -> .wav
assert fr == SAMPLING_RATE, "SAMPLING_RATE is not what is found in the wav file"
_, fbank, _ = tbmfcc(snd,
nwin=HAMMING_SIZE / 1000. *
SAMPLING_RATE,
nfft=2048,
fs=SAMPLING_RATE,
nceps=13)
x_fbank = np.append(x_fbank, fbank, axis=0)
assert t.getall().shape[0] == fbank.shape[
0], "MFCC and filterbank not of the same length (not on the same sampling rate)"
else:
fbank = None
with open(line.strip('"')[:-4] + '_fbanks.npy') as fbanksf:
fbank = np.load(fbanksf)
if fbank is not None:
# it seems filterbanks obtained with spectral are a little longer at the end
if DEBUG:
print("cutting the last",
fbank.shape[0] - t.getall().shape[0],
"frames from the filterbank")
fbank = fbank[:t.getall().shape[0]]
x_fbank = np.append(x_fbank, fbank, axis=0)
assert t.getall().shape[0] == fbank.shape[
0], "MFCC and filterbank not of the same length (not on the same sampling rate)"
if do_gammatones:
# load the wav sound (with Brian)
sound = loadsound(line.strip('"')[:-3] +
'wav') # .lab -> .wav
# compute the gammatones (see Brian's doc)
bw = 10**(0.037 + 0.785 * log10(center_frequencies))
gammatone = ApproximateGammatone(sound,
center_frequencies,
bw,
order=3)
g = gammatone.process()
# subsample the gammatones at the same rate than the MFCC's
# (just for practicality so that they are aligned...)
n_samples = g.shape[0] * 1. / (t.getall().shape[0] + 1
) # TODO check "+1"
### # do the harmonic mean (nth root of the product of the terms)
### g_sub = subsample_apply_f(g, n_samples, lambda z: np.power(np.prod(z), 1./n_samples))
g_sub = subsample_apply_f(
g, n_samples, lambda z: np.sqrt(np.sum(np.square(z))))
# compute the delta and delta of the subsampled gammatones
gamma_speed_accel = compute_speed_and_accel(g_sub)
# append
tmp = gamma_speed_accel[:t.getall().shape[0]] # TODO check
if tmp.shape[0] != t.getall().shape[0]: # TODO remove
print(line)
print(tmp.shape)
print(t.getall().shape)
print(n_samples)
print(g.shape)
print("exiting because of the mismatch")
sys.exit(-1)
x_gamma = np.append(x_gamma, tmp, axis=0)
elif line[0].isdigit():
start, end, state = line.split()[:3]
start = (int(start) + 9999) / (MFCC_TIMESTEP * 10000) # htk
end = (int(end) + 9999) / (MFCC_TIMESTEP * 10000) # htk
for i in range(start, end):
tmp_len_x -= 1
y.append(state)
y_spkr.append(speaker_label)
assert (len(y) == x.shape[0])
assert (len(y_spkr) == x.shape[0])
rootname = mlf[:-4]
np.save(rootname + '_xdata.npy', x)
np.save(rootname + '_xfbank.npy', x_fbank)
if do_gammatones:
np.save(rootname + '_xgamma.npy', x_gamma)
yy = np.array(y)
yy_spkr = np.array(y_spkr)
np.save(rootname + '_ylabels.npy', yy)
np.save(rootname + '_yspeakers.npy', yy_spkr)
print("length x:", len(x), "length y:", len(y), "length y_spkr:",
len(y_spkr))
print("shape x:", x.shape, "shape yy:", yy.shape, "shape yy_spkr:",
yy_spkr.shape)
if TEST:
tx = np.load(rootname + '_xdata.npy')
tx_fbank = np.load(rootname + '_xfbank.npy')
if do_gammatones:
tx_gamma = np.load(rootname + '_xgamma.npy')
ty = np.load(rootname + '_ylabels.npy')
ty_spkr = np.load(rootname + '_yspeakers.npy')
if np.all(tx == x) and np.all(ty == yy) and np.all(ty_spkr == yy_spkr):
assert_allclose(
tx_fbank,
x_fbank,
err_msg="x_fbank and its serialized version are not allclose")
if do_gammatones:
assert_allclose(
tx_gamma,
x_gamma,
err_msg=
"x_gamma and its serialized version are not allclose")
print("SUCCESS: serialized and current in-memory arrays are equal")
sys.exit(0)
else:
print(
"ERROR: serialized and current X (MFCC) or Y in-memory arrays differ!"
)
print("x (MFCC):", np.all(tx == x))
print("y (labels):", np.all(ty == yy))
print("y (speakers):", np.all(ty_spkr == yy_spkr))
sys.exit(-1)
def evaluate_environment(params, i_global, simulation_name, outputfolder, i_target=0, rank=1, speaker='adult', n_vow=5, normalize=False):
folder = outputfolder
############### Sound generation
if output:
print 'simulating vocal tract'
wavFile = parToWave(params, speaker, simulation_name, verbose=output, rank=rank) # call parToWave to generate sound file
# wavFile = par_to_wav(params, speaker, simulation_name, verbose=output, rank=rank) # call parToWave to generate sound file
if output:
print 'wav file '+str(wavFile)+' produced'
sound = loadsound(wavFile) # load sound file for brian.hears processing
if output:
print 'sound loaded'
############### Audio processing
sound = correct_initial(sound) # call correct_initial to remove initial burst
sound_resampled = get_resampled(sound)
# call get_resampled to adapt generated sound to AN model
sound_extended = get_extended(sound_resampled)
# call get_extended to equalize duration of all sounds
sound_extended.save(wavFile) # save current sound as sound file
os.system('cp '+wavFile+' '+folder+'data/vowel_'+str(i_target)+'_'+str(rank)+'.wav')
if playback:
print 'playing back...'
sound_extended.play(sleep=True) # play back sound file
if output:
print 'sound acquired, preparing auditory processing'
out = drnl(sound_extended) # call drnl to get cochlear activation
############### Classifier evaluation
flow_name = 'data/current_auditory_system.flow'
flow_file = open(flow_name, 'r') # open classifier file
flow = cPickle.load(flow_file) # load classifier
flow_file.close() # close classifier file
sample_vote_unnormalized = flow(out) # evaluate trained output units' responses for current item
if normalize:
sample_vote = normalize_activity(sample_vote_unnormalized)
else:
sample_vote = sample_vote_unnormalized
mean_sample_vote = np.mean(sample_vote, axis=0)
# average each output neurons' response over time
confidences = get_confidences(mean_sample_vote)
plot_reservoir_states(flow, sample_vote, i_target, folder, n_vow, rank)
return confidences
def main(args):
"""Main script."""
vowel = args["<vowel>"]
n_samples = int(args["--n_samples"])
n_channels = int(args["--n_channels"])
sigma = float(args["--sigma"])
uncompressed = args["--uncompressed"]
infant = args["--infant"]
monotone = args["--monotone"]
print 'generating ' + vowel + ' samples, infant mode: ' + str(infant)
np.random.seed() # numpy random seed w.r.t. global runtime
if infant:
speaker = 'infant'
else:
speaker = 'adult'
infant = True if speaker == 'infant' else False
initial_params_r = get_initial_params_r(vowel, infant=infant)
for i_global in xrange(n_samples):
folder = 'data/temp/' + vowel + '/'
if not os.path.isdir(folder):
os.makedirs(folder)
name = folder + vowel + '_' + str(i_global)
filename_act = name + '.dat.gz'
filename_wav = name + '.wav' # declare sound file name of current simulation
invalid = True
while invalid:
noise = np.random.randn(
16) * sigma # standard normally distributed vector
x = initial_params_r + noise # add mutation, Eq. 37
invalid = (x < 0.0).any() or (x > 1.0).any()
if invalid:
print 'sample rejected. resampling.'
params_tot = get_abs_coord(x, infant=infant)
############### Sound generation
wav_file = par_to_wav(params_tot,
speaker=speaker,
simulation_name=vowel,
different_folder=filename_wav,
monotone=monotone)
# call gesToWave to generate sound file
print 'wav file ' + str(wav_file) + ' produced'
sound = loadsound(
wav_file) # load sound file for brian.hears processing
print 'sound loaded'
sound = correct_initial(
sound) # call correct_initial to remove initial burst
sound_resampled = get_resampled(sound)
# call get_resampled to adapt generated sound to AN model
sound_extended = get_extended(sound_resampled)
# call get_extended to equalize duration of all sounds
sound_extended.save(wav_file) # save current sound as sound file
print 'sound acquired, preparing sound processing'
############### Audio processing
# call drnl to get cochlear activation
out = drnl(sound_extended, n_channels, uncompressed)
print 'writing auditory nerve response'
# create and open new output file in gzip write mode
with gzip.open(filename_act, 'wb') as outputfile:
out.dump(outputfile) # dump numpy array into output file
print 'done'
def process(
folder,
debug=False,
htk_mfc=False,
forcemfcext=False,
stereo_wav=False,
gammatones=False,
spectrograms=False,
filterbanks=False,
sox=True,
):
""" applies to all *.wav in folder """
# first find if we produce normalized MFCC, otherwise note it in the ext
# because we can then normalize on the whole corpus with another py script
mfc_extension = ".mfc_unnorm"
wcfg = open("wav_config", "r")
for line in wcfg:
if "ENORMALISE" in line:
mfc_extension = ".mfc"
if forcemfcext:
mfc_extension = ".mfc"
print "MFC extension:", mfc_extension
if gammatones:
try:
from brian import Hz, kHz
from brian.hears import loadsound, erbspace, Gammatone
except ImportError:
print >> sys.stderr, "You need Brian Hears"
print >> sys.stderr, "http://www.briansimulator.org/docs/\
hears.html"
sys.exit(-1)
if spectrograms:
try:
from pylab import specgram
except ImportError:
print >> sys.stderr, "You need Pylab"
sys.exit(-1)
fbanks = None
if filterbanks:
try:
sys.path.append("../spectral")
from spectral import Mel
except ImportError:
print >> sys.stderr, "You need spectral (in the parent folder)"
print >> sys.stderr, "https://github.com/mwv/spectral"
sys.exit(-1)
# run through all the folders and files in the path "folder"
# and put a header to the waves, save the originals as .rawaudio
# use HCopy to produce MFCC files according to "wav_config" file
for bdir, _, files in os.walk(folder):
for fname in files:
if fname[-4:] != ".wav":
continue
rawfname = bdir + "/" + fname[:-4] + ".rawaudio"
wavfname = bdir + "/" + fname
tempfname = bdir + "/" + fname[:-4] + "_temp.wav"
# temp fname with .wav for sox
mfccfname = bdir + "/" + fname[:-4] + mfc_extension
if sox:
shutil.move(wavfname, tempfname)
call(["sox", tempfname, wavfname])
# w/o headers, sox uses extension
shutil.move(tempfname, rawfname)
if htk_mfc:
call(["HCopy", "-C", "wav_config", wavfname, mfccfname])
srate = 16000
srate, sound = wavfile.read(wavfname)
if stereo_wav and len(sound.shape) == 2: # in mono sound is a list
sound = sound[:, 0] + sound[:, 1]
# for stereo wav, sum both channels
if gammatones:
gammatonefname = bdir + "/" + fname[:-4] + "_gamma.npy"
tmp_snd = loadsound(wavfname)
gamma_cf = erbspace(20 * Hz, 20 * kHz, N_GAMMATONES_FILTERS)
gamma_fb = Gammatone(tmp_snd, gamma_cf)
with open(gammatonefname, "w") as o_f:
npsave(o_f, gamma_fb.process())
if spectrograms:
powerspec, _, _, _ = specgram(
sound, NFFT=int(srate * SPECGRAM_WINDOW), Fs=srate, noverlap=int(srate * SPECGRAM_OVERLAP)
) # TODO
specgramfname = bdir + "/" + fname[:-4] + "_specgram.npy"
with open(specgramfname, "w") as o_f:
npsave(o_f, powerspec.T)
if filterbanks:
# convert to Mel filterbanks
if fbanks == None: # assume parameters are fixed
fbanks = Mel(
nfilt=N_FBANKS, # nb of filters in mel bank
alpha=0.97, # pre-emphasis
fs=srate, # sampling rate
frate=FBANKS_RATE, # frame rate
wlen=FBANKS_WINDOW, # window length
nfft=1024, # length of dft
mel_deltas=False, # speed
mel_deltasdeltas=False, # acceleration
)
fbank = fbanks.transform(sound)[0] # first dimension is for
# deltas & deltasdeltas
fbanksfname = bdir + "/" + fname[:-4] + "_fbanks.npy"
with open(fbanksfname, "w") as o_f:
npsave(o_f, fbank)
# TODO wavelets scattergrams / scalograms
print "dealt with file", wavfname
def main(args):
"""Main script."""
vowel = args["<vowel>"]
n_samples = int(args["--n_samples"])
n_channels = int(args["--n_channels"])
sigma = float(args["--sigma"])
uncompressed = args["--uncompressed"]
infant = args["--infant"]
monotone = args["--monotone"]
print 'generating ' + vowel + ' samples, infant mode: ' + str(infant)
np.random.seed() # numpy random seed w.r.t. global runtime
if infant:
speaker = 'infant'
else:
speaker = 'adult'
infant = True if speaker == 'infant' else False
initial_params_r = get_initial_params_r(vowel, infant=infant)
for i_global in xrange(n_samples):
folder = 'data/temp/'+vowel+'/'
if not os.path.isdir(folder):
os.makedirs(folder)
name = folder + vowel + '_' + str(i_global)
filename_act = name+'.dat.gz'
filename_wav = name+'.wav' # declare sound file name of current simulation
invalid = True
while invalid:
noise = np.random.randn(16) * sigma # standard normally distributed vector
x = initial_params_r + noise # add mutation, Eq. 37
invalid = (x < 0.0).any() or (x > 1.0).any()
if invalid:
print 'sample rejected. resampling.'
params_tot = get_abs_coord(x, infant=infant)
############### Sound generation
wav_file = par_to_wav(params_tot, speaker=speaker,
simulation_name=vowel,
different_folder=filename_wav, monotone=monotone)
# call gesToWave to generate sound file
print 'wav file ' + str(wav_file) + ' produced'
sound = loadsound(wav_file) # load sound file for brian.hears processing
print 'sound loaded'
sound = correct_initial(sound) # call correct_initial to remove initial burst
sound_resampled = get_resampled(sound)
# call get_resampled to adapt generated sound to AN model
sound_extended = get_extended(sound_resampled)
# call get_extended to equalize duration of all sounds
sound_extended.save(wav_file) # save current sound as sound file
print 'sound acquired, preparing sound processing'
############### Audio processing
# call drnl to get cochlear activation
out = drnl(sound_extended, n_channels, uncompressed)
print 'writing auditory nerve response'
# create and open new output file in gzip write mode
with gzip.open(filename_act, 'wb') as outputfile:
out.dump(outputfile) # dump numpy array into output file
print 'done'
def evaluate_environment(params,
i_global,
simulation_name,
outputfolder,
i_target=0,
rank=1,
speaker='adult',
n_vow=5,
normalize=False):
folder = outputfolder
############### Sound generation
if output:
print 'simulating vocal tract'
wavFile = parToWave(params,
speaker,
simulation_name,
verbose=output,
rank=rank) # call parToWave to generate sound file
# wavFile = par_to_wav(params, speaker, simulation_name, verbose=output, rank=rank) # call parToWave to generate sound file
if output:
print 'wav file ' + str(wavFile) + ' produced'
sound = loadsound(wavFile) # load sound file for brian.hears processing
if output:
print 'sound loaded'
############### Audio processing
sound = correct_initial(
sound) # call correct_initial to remove initial burst
sound_resampled = get_resampled(sound)
# call get_resampled to adapt generated sound to AN model
sound_extended = get_extended(sound_resampled)
# call get_extended to equalize duration of all sounds
sound_extended.save(wavFile) # save current sound as sound file
os.system('cp ' + wavFile + ' ' + folder + 'data/vowel_' + str(i_target) +
'_' + str(rank) + '.wav')
if playback:
print 'playing back...'
sound_extended.play(sleep=True) # play back sound file
if output:
print 'sound acquired, preparing auditory processing'
out = drnl(sound_extended) # call drnl to get cochlear activation
############### Classifier evaluation
flow_name = 'data/current_auditory_system.flow'
flow_file = open(flow_name, 'r') # open classifier file
flow = cPickle.load(flow_file) # load classifier
flow_file.close() # close classifier file
sample_vote_unnormalized = flow(
out) # evaluate trained output units' responses for current item
if normalize:
sample_vote = normalize_activity(sample_vote_unnormalized)
else:
sample_vote = sample_vote_unnormalized
mean_sample_vote = np.mean(sample_vote, axis=0)
# average each output neurons' response over time
confidences = get_confidences(mean_sample_vote)
plot_reservoir_states(flow, sample_vote, i_target, folder, n_vow, rank)
return confidences
def extract_from_mlf(mlf, do_gammatones):
x = np.ndarray((0, N_MFCC_COEFFS), dtype='float32')
x_fbank = np.ndarray((0, N_FILTERBANK_COEFFS), dtype='float32')
x_gamma = np.ndarray((0, N_GAMMATONES*3), dtype='float32')
y = []
y_spkr = []
with open(mlf) as f:
tmp_len_x = 0 # verify sizes
len_x = 0
end = 0
speaker_label = ''
for line in f:
line = line.rstrip('\n')
if len(line) < 1:
continue
if line[0] == '"':
assert tmp_len_x == 0, "the file above this one %s was mismatching x (%d frames) and y (%d frames) lengths by %d" % (line,
len_x, end, tmp_len_x)
speaker_label = line.split('/')[-2]
# load HTK's MFCC
t = htkmfc.open(line.strip('"')[:-3] + 'mfc') # .lab -> .mfc
x = np.append(x, t.getall(), axis=0)
len_x = t.getall().shape[0]
tmp_len_x = len_x
if TALKBOX_FBANKS: # do our own filterbanks TODO
fr, snd = wavfile.read(line.strip('"')[:-3] + 'wav') # .lab -> .wav
assert fr == SAMPLING_RATE, "SAMPLING_RATE is not what is found in the wav file"
_, fbank, _ = tbmfcc(snd, nwin=HAMMING_SIZE/1000.*SAMPLING_RATE, nfft=2048, fs=SAMPLING_RATE, nceps=13)
x_fbank = np.append(x_fbank, fbank, axis=0)
assert t.getall().shape[0] == fbank.shape[0], "MFCC and filterbank not of the same length (not on the same sampling rate)"
else:
fbank = None
with open(line.strip('"')[:-4] + '_fbanks.npy') as fbanksf:
fbank = np.load(fbanksf)
if fbank != None:
# it seems filterbanks obtained with spectral are a little longer at the end
if DEBUG:
print "cutting the last", fbank.shape[0] - t.getall().shape[0], "frames from the filterbank"
fbank = fbank[:t.getall().shape[0]]
x_fbank = np.append(x_fbank, fbank, axis=0)
assert t.getall().shape[0] == fbank.shape[0], "MFCC and filterbank not of the same length (not on the same sampling rate)"
if do_gammatones:
# load the wav sound (with Brian)
sound = loadsound(line.strip('"')[:-3] + 'wav') # .lab -> .wav
# compute the gammatones (see Brian's doc)
bw = 10**(0.037+0.785*log10(center_frequencies))
gammatone = ApproximateGammatone(sound, center_frequencies,
bw, order=3)
g = gammatone.process()
# subsample the gammatones at the same rate than the MFCC's
# (just for practicality so that they are aligned...)
n_samples = g.shape[0]*1./(t.getall().shape[0] + 1) # TODO check "+1"
### # do the harmonic mean (nth root of the product of the terms)
### g_sub = subsample_apply_f(g, n_samples, lambda z: np.power(np.prod(z), 1./n_samples))
g_sub = subsample_apply_f(g, n_samples, lambda z: np.sqrt(np.sum(np.square(z))))
# compute the delta and delta of the subsampled gammatones
gamma_speed_accel = compute_speed_and_accel(g_sub)
# append
tmp = gamma_speed_accel[:t.getall().shape[0]] # TODO check
if tmp.shape[0] != t.getall().shape[0]: # TODO remove
print line
print tmp.shape
print t.getall().shape
print n_samples
print g.shape
print "exiting because of the mismatch"
sys.exit(-1)
x_gamma = np.append(x_gamma, tmp, axis=0)
elif line[0].isdigit():
start, end, state = line.split()[:3]
start = (int(start)+9999)/(MFCC_TIMESTEP * 10000) # htk
end = (int(end)+9999)/(MFCC_TIMESTEP * 10000) # htk
for i in xrange(start, end):
tmp_len_x -= 1
y.append(state)
y_spkr.append(speaker_label)
assert(len(y) == x.shape[0])
assert(len(y_spkr) == x.shape[0])
rootname = mlf[:-4]
np.save(rootname + '_xdata.npy', x)
np.save(rootname + '_xfbank.npy', x_fbank)
if do_gammatones:
np.save(rootname + '_xgamma.npy', x_gamma)
yy = np.array(y)
yy_spkr = np.array(y_spkr)
np.save(rootname + '_ylabels.npy', yy)
np.save(rootname + '_yspeakers.npy', yy_spkr)
print "length x:", len(x), "length y:", len(y), "length y_spkr:", len(y_spkr)
print "shape x:", x.shape, "shape yy:", yy.shape, "shape yy_spkr:", yy_spkr.shape
if TEST:
tx = np.load(rootname + '_xdata.npy')
tx_fbank = np.load(rootname + '_xfbank.npy')
if do_gammatones:
tx_gamma = np.load(rootname + '_xgamma.npy')
ty = np.load(rootname + '_ylabels.npy')
ty_spkr = np.load(rootname + '_yspeakers.npy')
if np.all(tx==x) and np.all(ty==yy) and np.all(ty_spkr==yy_spkr):
assert_allclose(tx_fbank, x_fbank, err_msg="x_fbank and its serialized version are not allclose")
if do_gammatones:
assert_allclose(tx_gamma, x_gamma, err_msg="x_gamma and its serialized version are not allclose")
print "SUCCESS: serialized and current in-memory arrays are equal"
sys.exit(0)
else:
print "ERROR: serialized and current X (MFCC) or Y in-memory arrays differ!"
print "x (MFCC):", np.all(tx==x)
print "y (labels):", np.all(ty==yy)
print "y (speakers):", np.all(ty_spkr==yy_spkr)
sys.exit(-1)