ps = {}
with open(dictfn2, 'w') as dictf:
dictf.write('!SIL SIL\n')
# FIXME: re-enable once we have noise tokens in our transcripts
# dictf.write('<SPOKEN_NOISE> SPN\n')
dictf.write('<UNK> SPN\n')
# dictf.write('<NOISE> NSN\n')
for token in sorted(utt_dict):
ipa = utt_dict[token]
xsr = ipa2xsampa(token, ipa, spaces=True)
xs = xsr.replace('-', '').replace('\' ', '\'').replace(' ', ' ')
dictf.write((u'%s %s\n' % (token, xs)).encode('utf8'))
for p in xs.split(' '):
if len(p) < 1:
print u"****ERROR: empty phoneme in : '%s' ('%s', ipa: '%s')" % (
xs, xsr, ipa)
pws = p[1:] if p[0] == '\'' else p
if not pws in ps:
ps[pws] = set([p])
def lex_edit(token):
global lex, lex_token, lex_entry, lex_base
lex_set_token(token)
while True:
lex_paint_main()
c = stdscr.getch()
if c == ord('q'):
lex.save()
break
# generate de-mary
elif c == ord('g'):
tts.set_locale('de')
tts.set_engine('mary')
tts.set_voice('bits3')
ipas = tts.gen_ipa(lex_base)
tts.say_ipa(ipas)
lex_entry['ipa'] = ipas
# generate de-espeak
elif c == ord('h'):
tts.set_locale('de')
tts.set_engine('espeak')
tts.set_voice('de')
ipas = tts.gen_ipa(lex_base)
tts.set_locale('de')
tts.set_engine('mary')
tts.set_voice('bits3')
tts.say_ipa(ipas)
lex_entry['ipa'] = ipas
# generate en-mary
elif c == ord('l'):
tts.set_locale('en-US')
tts.set_engine('mary')
tts.set_voice('cmu-rms-hsmm')
ipas = tts.gen_ipa(lex_base)
tts.say_ipa(ipas)
lex_entry['ipa'] = ipas
# generate fr-mary
elif c == ord('k'):
tts.set_locale('fr')
tts.set_engine('mary')
tts.set_voice('pierre-voice-hsmm')
ipas = tts.gen_ipa(lex_base)
tts.say_ipa(ipas)
lex_entry['ipa'] = ipas
# generate de-sequitur
elif c == ord('j'):
ipas = sequitur_gen_ipa(lex_base)
tts.set_locale('de')
tts.set_engine('mary')
tts.set_voice('bits3')
tts.say_ipa(ipas)
lex_entry['ipa'] = ipas
# speak de mary unitsel
elif c == ord('p'):
if len(lex_entry['ipa']) == 0:
continue
ipas = lex_entry['ipa']
tts.set_locale('de')
tts.set_engine('mary')
tts.set_voice('bits3')
tts.say_ipa(ipas)
# speak de mary hsmm
elif c == ord('o'):
if len(lex_entry['ipa']) == 0:
continue
ipas = lex_entry['ipa']
tts.set_locale('de')
tts.set_engine('mary')
tts.set_voice('dfki-pavoque-neutral-hsmm')
tts.say_ipa(ipas)
# speak fr mary hsmm
elif c == ord('i'):
if len(lex_entry['ipa']) == 0:
continue
ipas = lex_entry['ipa']
tts.set_locale('fr')
tts.set_engine('mary')
tts.set_voice('pierre-voice-hsmm')
tts.say_ipa(ipas)
# speak en mary hsmm
elif c == ord('u'):
ipas = lex_entry['ipa']
tts.set_locale('en-US')
tts.set_engine('mary')
tts.set_voice('cmu-rms-hsmm')
tts.say_ipa(ipas)
# edit token
elif c == ord('t'):
token = utils.edit_popup(stdscr, ' Token ', '')
lex_set_token(token)
# edit XS
elif c == ord('e'):
ipas = lex_entry['ipa']
xs = ipa2xsampa(lex_token, ipas, stress_to_vowels=False)
xs = utils.edit_popup(stdscr, ' X-SAMPA ', xs)
try:
ipas = xsampa2ipa(lex_token, xs)
lex_entry['ipa'] = ipas
except:
pass
lex = Lexicon()
print "loading lexicon...done."
#
# export
#
utils.mkdirs(WORKDIR)
with codecs.open('%s/train.lex' % WORKDIR, 'w', 'utf8') as trainf, \
codecs.open('%s/test.lex' % WORKDIR, 'w', 'utf8') as testf, \
codecs.open('%s/all.lex' % WORKDIR, 'w', 'utf8') as allf :
cnt = 0
for word in lex:
ipa = lex[word]['ipa']
xs = ipa2xsampa (word, ipa, spaces=True, stress_to_vowels=False)
if cnt % 10 == 0:
testf.write (u'%s %s\n' % (word, xs))
else:
trainf.write (u'%s %s\n' % (word, xs))
allf.write (u'%s %s\n' % (word, xs))
cnt += 1
allf.write (u'<s> %s </s>\n' % ' '.join(tokens))
#
# wlist, dictionary
#
with codecs.open('%s/hal.vocab' % WORKDIR, 'w', 'utf8') as vocabf, \
codecs.open('%s/hal.dic' % WORKDIR, 'w', 'utf8') as dicf:
vocabf.write('</s>\n')
vocabf.write('<s>\n')
for token in sorted(all_tokens):
vocabf.write(u'%s\n' % token)
ipa = lex[token]['ipa']
xs = ipa2xsampa(token, ipa)
xa = xsampa2xarpabet(token, xs)
dicf.write(u'%s %s\n' % (token, xa))
#
# generate sphinx lm
#
os.system('text2idngram -vocab %s/hal.vocab -idngram %s/hal.idngram < %s/hal.txt' % (WORKDIR, WORKDIR, WORKDIR))
os.system('idngram2lm -calc_mem -vocab_type 0 -idngram %s/hal.idngram -vocab %s/hal.vocab -arpa %s/hal.arpa' % (WORKDIR, WORKDIR, WORKDIR))
os.system('sphinx_lm_convert -i %s/hal.arpa -o %s/hal.lm.DMP' % (WORKDIR, WORKDIR))
tts.say_ipa(ipas)
# edit token
elif c == ord('t'):
token = utils.edit_popup(stdscr, ' Token ', '')
lex_set_token (token)
# edit XS
elif c == ord('e'):
ipas = lex_entry['ipa']
xs = ipa2xsampa (lex_token, ipas, stress_to_vowels=False)
xs = utils.edit_popup(stdscr, ' X-SAMPA ', xs)
try:
ipas = xsampa2ipa (lex_token, xs)
lex_entry['ipa'] = ipas
except:
pass
#
# fini
def export_sphinx_case(work_dir, sphinxtrain_cfg_fn):
#
# language model
#
utils.mkdirs('%s' % work_dir)
fn = '%s/prompts.sent' % work_dir
with codecs.open(fn, 'w', 'utf8') as outf:
for cfn in ts_all:
transcript = transcripts[cfn]['ts']
outf.write('%s\n' % transcript)
print "%s written." % fn
print
fn = '%s/wlist.txt' % work_dir
with codecs.open(fn, 'w', 'utf8') as outf:
for word in lex:
outf.write('%s\n' % word)
print "%s written." % fn
print
#
# create work_dir structure
#
mfcc_dir = "%s/mfcc" % work_dir
utils.mkdirs('%s/logs' % work_dir)
utils.mkdirs('%s/etc' % work_dir)
utils.mkdirs('%s' % mfcc_dir)
# generate sphinx_train.cfg, featdir in there
# inf = codecs.open ('data/src/speech/sphinx_train.cfg', 'r', 'utf8')
# outf = codecs.open ('%s/etc/sphinx_train.cfg' % work_dir, 'w', 'utf8')
# for line in inf:
# s = line.replace('%FEATDIR%', mfcc_dir).replace('%WORKDIR%', work_dir)
# outf.write (s)
# inf.close()
# outf.close()
utils.copy_file(sphinxtrain_cfg_fn, '%s/etc/sphinx_train.cfg' % work_dir)
utils.copy_file('data/src/speech/sphinx-voxforge.filler',
'%s/etc/voxforge.filler' % work_dir)
utils.copy_file('data/src/speech/sphinx-feat.params',
'%s/etc/feat.params' % work_dir)
# generate dict
phoneset = set()
pdfn = '%s/etc/voxforge.dic' % work_dir
with codecs.open(pdfn, 'w', 'utf8') as pdf:
for word in lex:
ipa = lex[word]['ipa']
xs = ipa2xsampa(word, ipa)
xa = xsampa2xarpabet(word, xs)
pdf.write(u'%s %s\n' % (word, xa))
phones = xa.split(' ')
for phone in phones:
if len(phone.strip()) == 0:
print u"***ERROR: empty phone detected in lex entry %s %s" % (
word, ipa)
phoneset.add(phone)
print "%s written." % pdfn
print
print "Got %d phones." % len(phoneset)
phfn = '%s/etc/voxforge.phone' % work_dir
with codecs.open(phfn, 'w', 'utf8') as phf:
for phone in phoneset:
phf.write(u'%s\n' % phone)
phf.write(u'SIL\n')
print "%s written." % phfn
print
#
# prompts
#
train_fifn = '%s/etc/voxforge_train.fileids' % work_dir
train_tsfn = '%s/etc/voxforge_train.transcription' % work_dir
test_fifn = '%s/etc/voxforge_test.fileids' % work_dir
test_tsfn = '%s/etc/voxforge_test.transcription' % work_dir
runfeatfn = '%s/run-feat.sh' % work_dir
SPHINXFE = "sphinx_fe -i '%s' -part 1 -npart 1 -ei wav -o '%s' -eo mfc -nist no -raw no -mswav yes -samprate 16000 -lowerf 130 -upperf 6800 -nfilt 25 -transform dct -lifter 22 >>logs/mfcc%02d.log 2>&1 &\n"
with codecs.open(runfeatfn, 'w', 'utf8') as runfeatf:
runfeatf.write('#!/bin/bash\n\n')
cnt = 0
for cfn in ts_all:
w16filename = "%s/%s.wav" % (wav16_dir, cfn)
mfcfilename = "mfcc/%s.mfc" % cfn
runfeatf.write(SPHINXFE % (w16filename, mfcfilename, cnt))
cnt = (cnt + 1) % NJOBS
if cnt == 0:
runfeatf.write('wait\n')
print "%s written." % runfeatfn
with codecs.open (train_fifn, 'w', 'utf8') as train_fif, \
codecs.open (train_tsfn, 'w', 'utf8') as train_tsf, \
codecs.open (test_fifn, 'w', 'utf8') as test_fif, \
codecs.open (test_tsfn, 'w', 'utf8') as test_tsf:
for cfn in ts_train:
train_fif.write('%s\n' % cfn)
train_tsf.write(u'<s> %s </s> (%s)\n' % (ts_train[cfn]['ts'], cfn))
for cfn in ts_test:
test_fif.write('%s\n' % cfn)
test_tsf.write(u'<s> %s </s> (%s)\n' % (ts_test[cfn]['ts'], cfn))
print "%s written." % train_tsfn
print "%s written." % train_fifn
print "%s written." % test_tsfn
print "%s written." % test_fifn
utils.copy_file('data/src/speech/sphinx-run.sh',
'%s/sphinx-run.sh' % work_dir)