def readNonEmptyTokensTextGrid(annotationURI, whichLevel, initialTimeOffset=0):
'''
######################
# prepare list of phrases from ANNOTATION. remove empy annotaion tokens
'''
try:
annotationTokenListA = TextGrid2WordList(annotationURI, whichLevel)
except Exception as errorMsg:
sys.exit(str(errorMsg))
for currAnnoTsAndToken in annotationTokenListA:
currAnnoTsAndToken[0] = float(currAnnoTsAndToken[0])
currAnnoTsAndToken[0] += initialTimeOffset
currAnnoTsAndToken[1] = float(currAnnoTsAndToken[1])
currAnnoTsAndToken[1] += initialTimeOffset
# store to file .anno
baseN = os.path.basename(annotationURI)
dir = os.path.dirname(annotationURI)
annotationURI_anno = os.path.join(dir,baseN+'.anno')
writeListOfListToTextFile(annotationTokenListA, None, annotationURI_anno )
annotationTokenListNoPauses = []
#########
# remove empty phrases
for currAnnoTsAndToken in annotationTokenListA:
if currAnnoTsAndToken[2] != "" and not (currAnnoTsAndToken[2].isspace()): # skip empty phrases
annotationTokenListNoPauses.append(currAnnoTsAndToken)
return annotationTokenListA, annotationTokenListNoPauses
def _constructLogLiksTable(self, usePersistentProbs):
PATH_LOOKUP_DUR_TABLE = PATH_LOGS + '/lookupTable'
logger.info("path lookup table: " + PATH_LOOKUP_DUR_TABLE)
if usePersistentProbs and os.path.exists(PATH_LOOKUP_DUR_TABLE):
self.lookupTableLogLiks = numpy.loadtxt(PATH_LOOKUP_DUR_TABLE)
logger.info("reading lookup table from {}".format( PATH_LOOKUP_DUR_TABLE ))
# if table covers max dur
if self.lookupTableLogLiks.shape[0] >= self.R_MAX:
return
else:
self.lookupTableLogLiks = numpy.empty((self.R_MAX, self.MAX_ALLOWED_DURATION_RATIO * self.R_MAX + 1))
self.lookupTableLogLiks.fill(-Infinity)
# otherwise construct
logging.info("constructing duration Probability lookup Table...")
quantileVals = linspace(self.minVal, self.maxVal, self.numDurs )
liks = numpy.zeros((self.numDurs,1))
for d in range(0,self.numDurs):
liks[d] = norm.pdf(quantileVals[d])
for currMaxDur in range(1,int(self.R_MAX)+1):
self._constructLogLikDistrib( currMaxDur, liks)
writeListOfListToTextFile(self.lookupTableLogLiks, None , PATH_LOOKUP_DUR_TABLE)
def _createWordMLFandDict(self):
#txtTur to METU. txtMETU as persistent file not really needed. Kept only for reference
baseNameAudioFile = os.path.splitext(self.pathToAudioFile)[0]
METUBETfileName = baseNameAudioFile + LYRICS_TXT_METUBET_EXT
if (self.loadLyricsFromFile == 1):
METULyrics = PhonetizerOld.turkishScriptLyrics2METUScriptLyricsFile(baseNameAudioFile + LYRICS_TXT_EXT, METUBETfileName)
else:
# TODO: change this step
METULyrics = PhonetizerOld.turkishScriptLyrics2METUScriptLyrics(self.lyrics, METUBETfileName)
# create Word-level mlf:
baneN = os.path.basename(self.pathToAudioFile)
baneN = os.path.splitext(baneN)[0]
headerLine = baneN + ' ' + METULyrics
writeListOfListToTextFile([], headerLine, '/tmp/prompts')
# prompts2mlf
mlfName = '/tmp/tmp' + HTK_MLF_WORD_ANNO_SUFFIX
prompts2mlf = os.path.abspath('prompts2mlf')
pipe = subprocess.Popen(['/usr/bin/perl', prompts2mlf, mlfName, '/tmp/prompts'])
pipe.wait()
# phonetize
dictName = '/tmp/lexicon2'
PhonetizerOld.METULyrics2phoneticDict(METUBETfileName, dictName, self.withSynthesis)
return (dictName, mlfName, METULyrics )
def _mapB(self,observations):
'''
Required implementation for _mapB. Refer to _BaseHMM for more details.
This method highly optimizes the running time, since all PDF calculations
are done here once in each training iteration.
- self.Bmix_map - computesand maps Bjm(Ot) to Bjm(t).
'''
self.B_map = numpy.zeros( (self.n,len(observations)), dtype=self.precision)
if self.usePersistentFiles and os.path.exists(self.PATH_BMAP):
self.B_map = numpy.loadtxt(self.PATH_BMAP)
if self.B_map.shape[1] == len(observations):
# sys.exit('{} does not store all feature vectors. delete it and generate them again'.format(self.PATH_BMAP))
return
self.Bmix_map = numpy.zeros( (self.n,self.m,len(observations)), dtype=self.precision)
for j in xrange(self.n):
for t in xrange(len(observations)):
lik = self._calcbjt(j, t, observations[t])
if lik == 0:
logging.warning("obs likelihood at time {} for state {} = 0. Repair by adding {}".format(t,j, MINIMAL_PROB))
lik = MINIMAL_PROB
self.B_map[j][t] = lik
self._normalizeBByMax()
# normalize over states
for t in xrange(len(observations)):
self.B_map[:,t] = _normalize(self.B_map[:,t])
logging.debug("sum={} at time {}".format(sum(self.B_map[:,t]), t))
if self.usePersistentFiles:
writeListOfListToTextFile(self.B_map, None , self.PATH_BMAP)
def _initKappas(self, lenObservations):
'''
kappas[t][s] - starting and staying at time t in same currState s.
WITH LogLik
'''
if lenObservations <= self.R_MAX:
sys.exit("observations are only {}, R_max = {}. not able to run initialization. Increase size of observations".format(lenObservations,self.R_MAX))
print 'init kappas...'
self.kappas = numpy.empty((self.R_MAX,self.n), dtype=self.precision)
# if some kappa[t, state] = -INFINITY and phi[t,state] = -INFINITY, no initialization is possilbe (e.g. not possible to choose max btw kappa and phi)
self.kappas.fill(numpy.log(MINIMAL_PROB))
for currState in range(self.n):
sumObsProb = 0
currRefMax = self.durationPdf.getMaxRefDur( self.durationMap[currState])
currLogPi = numpy.log(self.pi[currState])
for t in range(1,int(currRefMax)+1):
updateQuantity, sumObsProb = self._calcUpdateQuantity(t-1, t, currState, 0, sumObsProb)
self.kappas[t-1,currState] = currLogPi + updateQuantity
#sanity check. for debug
if self.kappas[t-1,currState] == 0:
print "underflow error at time {}, currState {}".format(t-1, currState)
writeListOfListToTextFile(self.kappas, None , PATH_LOGS + '/kappas')
def _mapB_OLD(self, observations):
'''
Required implementation for _mapB. Refer to _BaseHMM for more details.
This method highly optimizes the running time, since all PDF calculations
are done here once in each training iteration.
- self.Bmix_map - computesand maps Bjm(Ot) to Bjm(t).
log precomputed
'''
# return
if self.usePersistentFiles and os.path.exists(self.PATH_BMAP):
self.logger.info("loading probs all observations from {}".format(self.PATH_BMAP))
self.B_map = numpy.loadtxt(self.PATH_BMAP)
# check length
if self.B_map.shape[1] == len(observations) and self.B_map.shape[0] == self.n:
# sys.exit('{} does not store all feature vectors. delete it and generate them again'.format(self.PATH_BMAP))
self.B_map = numpy.log( self.B_map)
return
else:
self.logger.info("file {} found, but has not the expected num of states {} or observations {}".format(self.PATH_BMAP, self.n, len(observations)) )
self.B_map = numpy.zeros( (self.n,len(observations)), dtype=self.precision)
self.Bmix_map = numpy.zeros( (self.n,self.m,len(observations)), dtype=self.precision)
for j in xrange(self.n):
for t in xrange(len(observations)):
self.logger.debug("at calcbjt at state {} and time {}...\n".format(j, t))
lik = self._calcbjt(j, t, observations[t])
if lik == 0:
self.logger.debug("obs likelihood at time {} for state {} = 0. Repair by adding {}".format(t,j, MINIMAL_PROB))
lik = MINIMAL_PROB
self.B_map[j,t] = lik
# normalize over states
for t in xrange(len(observations)):
self.B_map[:,t] = _normalize(self.B_map[:,t])
self.logger.debug("sum={} at time {}".format(sum(self.B_map[:,t]), t))
if self.usePersistentFiles:
writeListOfListToTextFile(self.B_map, None , self.PATH_BMAP)
self.B_map = numpy.log( self.B_map)
def tokenList2TabFile( listTsAndPhonemes, baseNameAudioFile, whichSuffix):
'''
convenience method.
'''
# timeshift
# for index in range(len(listTsAndPhonemes)):
# listTsAndPhonemes[index][0] = listTsAndPhonemes[index][0] + timeShift
# if (len(listTsAndPhonemes[index]) == 3):
# del listTsAndPhonemes[index][1]
phonemeAlignedfileName = baseNameAudioFile + whichSuffix
writeListOfListToTextFile(listTsAndPhonemes, 'startTs endTs phonemeOrWord\n', phonemeAlignedfileName)
logging.debug('phoneme level alignment written to file: ', phonemeAlignedfileName)
return phonemeAlignedfileName
def _viterbiForcedDur(self, observations):
# sanity check. make sure durations are init from score
print "decoding..."
for t in range(self.R_MAX,len(observations)):
for currState in xrange(1, self.n):
self.computePhi(t, currState) # get max duration quantities
# for t in range(self.R_MAX+19, self.R_MAX+20):
# for currState in xrange(57, 59):
# self.computePhi(t, currState) # get max duration quantities
writeListOfListToTextFile(self.phi, None , PATH_LOGS + '/phi')
# return for backtracking
return self.chi, self.psi
def _initBeginingPhis(self, lenObservations):
'''
init phis when t < self.R_MAX
'''
self._initKappas(lenObservations)
# for convenience put as class vars
self.phi = numpy.empty((lenObservations,self.n),dtype=self.precision)
self.phi.fill(-Infinity)
# self.phi = numpy.loadtxt(PATH_LOGS + '/phi_init', dtype=self.precision)
# return
# init t=0
# for currState in range(self.n): self.phi[0,currState] = self.kappas[currState,0]
self.phi[0,:] = self.kappas[0,:]
# init first state = kappa (done to allow self.getMaxPhi_slow to access prev. currState)
self.phi[:len(self.kappas[:,0]),0] = self.kappas[:,0]
# select bigger (kappa and phi_star)
for t in range(1,int(self.R_MAX)):
self.logger.debug("at time t={}".format(t) )
# phi start makes sence only from second state
for currState in range(1, self.n):
phiStar, fromState, maxDurIndex = self.computePhiStar(t, currState)
# take bigger : eq:deltaStarOrKappa
if phiStar > self.kappas[t,currState] :
self.phi[t,currState] = phiStar
self.psi[t,currState] = fromState
self.chi[t,currState] = maxDurIndex
else:
self.logger.debug( " kappa more than phi at time {} and state {}".format(t, currState))
self.phi[t, currState] = self.kappas[t, currState]
# kappas mean still at beginning state
self.psi[t,currState] = currState
self.chi[t,currState] = t
writeListOfListToTextFile(self.phi, None , PATH_LOGS + '/phi_init')
def _mlf2PraatFormat( listTsAndPhonemes, timeShift, baneNameAudioFile, whichSuffix):
# timeshift
for index in range(len(listTsAndPhonemes)):
listTsAndPhonemes[index][0] = listTsAndPhonemes[index][0] + timeShift
if (len(listTsAndPhonemes[index]) == 3):
del listTsAndPhonemes[index][1]
phonemeAlignedfileName = baneNameAudioFile + whichSuffix
writeListOfListToTextFile(listTsAndPhonemes, 'startTs phonemeOrWord\n', phonemeAlignedfileName)
logger.debug('phoneme level alignment written to file: ', phonemeAlignedfileName)
return phonemeAlignedfileName
'''
def getReferenceDurations(URI_recording_noExt, decoder, evalLevel):
'''
timestamps of words according to reference durations read from score. Used to obtain so called 'score-deviation' metric
not used in decoding
'''
annotationURI = URI_recording_noExt + ANNOTATION_EXT
##### get duration of initial silence
try:
annotationTokenListA = TextGrid2WordList(annotationURI, evalLevel)
# just copy duration of silence in groundTruth
annoTsAndToken = annotationTokenListA[0]
if annoTsAndToken[2] != "" and not(annoTsAndToken[2].isspace()): # skip empty phrases
logger.warn("annotaiton {} starts with non-sil token ".format(annotationURI))
finalSilFram = float(annoTsAndToken[0]) * NUM_FRAMES_PERSECOND
else:
finalSilFram = float(annoTsAndToken[1]) * NUM_FRAMES_PERSECOND
except :
# if no Gr Truth annotation file (or needed layer) present - take from model
finalSilFram = 0
countFirstStateFirstWord = decoder.lyricsWithModels.listWords[0].syllables[0].phonemes[0].numFirstState
for i in range(countFirstStateFirstWord):
finalSilFram += decoder.lyricsWithModels.statesNetwork[i].getDurationInFrames()
grTruthWordList = expandlyrics2Words (decoder.lyricsWithModels, decoder.lyricsWithModels.statesNetwork, finalSilFram, _constructTimeStampsForWord)
grTruthDurationfileExtension = '.grTruthDuration'
writeListOfListToTextFile(grTruthWordList, None , URI_recording_noExt + grTruthDurationfileExtension )
# TODO: could be done easier with this code, and check last method in Word
# grTruthWordList = testT(decoder.lyricsWithModels)
return grTruthWordList
def testReadListOfListTextFile_gen():
URIfile = '/Users/joro/Downloads/kimseye-annotation-score-to-audio.txt'
shiftedNakarat = []
inNakarat = 0
detectedTokenList = readListOfListTextFile_gen(URIfile)
# get TS
for entry in detectedTokenList:
if entry[3] == 'D5-NAKARAT-n1':
beginTs = entry[0]
if entry[3] == 'D5-NAKARAT*-n1':
endTs = entry[0]
break
endTs = 111.687981859
for entry in detectedTokenList:
if entry[3] == 'D5-NAKARAT-n1':
inNakarat = 1
if entry[3] == 'D5-NAKARAT*-n1':
inNakarat = 0
break
if inNakarat:
entry[0] += (endTs - beginTs)
shiftedNakarat.append(entry)
writeListOfListToTextFile(shiftedNakarat,None, 'shiftedNakarat.txt', toFlip=False)
print detectedTokenList