def guessStress(input_string):
b = BlickLoader()
vows = {x[:-1]: { y[-1] for y in b.vowels if y[:-1] == x[:-1]} for x in b.vowels}
phones = input_string.split(" ")
stress_pattern_space = [ vows[x] for x in phones if x in vows]
revised_string = []
for p in phones:
if p in vows:
revised_string.append(p+"%s")
else:
revised_string.append(p)
revised_string = ' '.join(revised_string)
possible_patterns = list(itertools.product(*stress_pattern_space))
possible_strings = [revised_string % x for x in possible_patterns]
best = 1000
pattern = ''
for p in possible_strings:
if p in word_list:
return p
score = b.assessWord(p)
if score < best:
best = score
pattern = p
return pattern
def getPhonScores(puns):
b = BlickLoader()
scoredPuns = {}
for pun in puns:
words = pun.split('#')
del words[len(words) - 1]
goodWords = []
for w in words:
if "D I C T" not in w:
goodWords.append(w)
scoredPuns[pun] = sum(b.assessWord(w.strip()) for w in goodWords)
return scoredPuns
def train_wakeword_model(audio_train_loader,
vocab_list,
label_model,
beam_size=3,
num_hypotheses=5,
query_by_string=False):
wakeword_model = {}
if query_by_string:
# load ww model produced by MFA from config
keywords = config["wakeword_model"]
# load blick
b = BlickLoader()
for i, _, y_hat in enumerate(keywords.items()):
w = b.assessWord(y_hat)
# for each keyword, append the tuple(hypotheses + weights) to the list
# only one hypothesis if using MFA
wakeword_model[i] = (y_hat, w)
else:
# train ww model from scratch
for i in audio_train_loader:
posteriors_i = label_model(i)
# decode using CTC, vocab_list is A (labels)
decoder = CTCBeamDecoder(self.vocab_list,
beam_width=self.beam_size,
blank_id=self.vocab_list.index('_'))
beam, beam_scores, _, _ = decoder.decode(posteriors_i)
for j in range(num_hypotheses):
y_hat = beam[j] # hypothesis
log_prob_post = beam_scores[j]
w = log_prob_post**-1
# for each keyword, append the tuple(hypotheses + weights) to the list
wakeword_model[i].append((y_hat, w))
return wakeword_model
def getPhonotacticProb(input_string,use_blick=True,no_stress=False):
if use_blick:
if no_stress:
b = BlickLoader(grammarType='NoStress')
else:
b = BlickLoader()
return b.assessWord(str(input_string))
if no_stress:
input_string = re.sub('[0-9]','',input_string)
SPprob = 0.0
BPprob = 0.0
phones = input_string.split(" ")
for i in range(len(phones)):
patt = [any_segment] * i
patt.append(phones[i])
pattern = '^'+' '.join(patt) +'.*$'
totPattern = '^'+' '.join([any_segment] * (i+1)) +'.*$'
if no_stress:
count = len(PhonoString.objects.filter(NoStress__regex = pattern))
totCount = len(PhonoString.objects.filter(NoStress__regex = totPattern))
else:
count = len(PhonoString.objects.filter(Transcription__regex = pattern))
totCount = len(PhonoString.objects.filter(Transcription__regex = totPattern))
SPprob += float(count) / float(totCount)
if i != len(phones)-1:
patt = [any_segment] * i
patt.extend([phones[i],phones[i+1]])
pattern = '^'+' '.join(patt) +'.*$'
totPattern = '^'+' '.join([any_segment] * (i+2)) +'.*$'
if no_stress:
count = len(PhonoString.objects.filter(NoStress__regex = pattern))
totCount = len(PhonoString.objects.filter(NoStress__regex = totPattern))
else:
count = len(PhonoString.objects.filter(Transcription__regex = pattern))
totCount = len(PhonoString.objects.filter(Transcription__regex = totPattern))
BPprob += float(count) / float(totCount)
SPprob = SPprob / float(len(phones))
BPprob = BPprob / float(len(phones)-1)
return (SPprob,BPprob)
if syl.nucleus in SHORT_VOWELS:
return False
if is_first and syl.onset and syl.onset[0] == "ZH":
return False
# if is_last and stress_lvl == 1 and len(syl.coda) == 0:
# return False
if syl.onset and syl.coda and syl.onset[0] == "S" and not syl.onset[
-1] == "T" and syl.nucleus in SHORT_VOWELS:
if syl.coda[0] == syl.onset[-1]:
return False
if stress_lvl != 1 and syl.nucleus not in {
"AH", "ER", "IH", "IY", "OW", "UW"
}:
return False
return True
blick_rater = BlickLoader()
words = []
for i in range(100):
word = getWord()
score, rules = blick_rater.assessWord(word.replace(" ", " "),
includeConstraints=True)
score = exp(-score)
words.append([word, score, rules])
for word, score, rules in sorted(words, key=itemgetter(1)):
if score > 0.00001:
print(word, score, rules)
print()
for word in open(dictionary):
word = word.strip()
if word != "" and not word.endswith("'s"):
chain.add(word.lower())
#Make a word, check if it is within set range, search google for it, save it. Up to 5000 words
words = 0
while words < 5000:
word = "".join(chain.random_output())
if len(word) > 4 and len(word) < 10:
score = 100
blickified = phonetify(word)
for blicked in blickified:
try:
#sometimes this bails out, instead of tracking it down each time, this was an easy out
thisscore = b.assessWord(blicked)
except:
score = 100
if thisscore < score:
score=thisscore
if score > 0 and score < 18:
try:
first_url = search('"' + word + '"',num=10, stop=1)
for x in range(1,10):
this_url = first_url.next()
print str(score) + ',' + word + " - bad " + this_url
except StopIteration:
with open('proop.out', 'a') as outfile:
outfile.write(str(score) + ',' + word + '\n')
print str(score) + ',' + word + " - good"
words += 1
