def copy_scripts_and_config_files(work_dir, kaldi_root):
misc.copy_file('data/src/speech/kaldi-run-lm.sh', '%s/run-lm.sh' % work_dir)
# misc.copy_file ('data/src/speech/kaldi-run-am.sh', '%s/run-am.sh' % work_dir)
# misc.copy_file ('data/src/speech/kaldi-run-nnet3.sh', '%s/run-nnet3.sh' % work_dir)
misc.copy_file('data/src/speech/kaldi-run-chain.sh',
'%s/run-chain.sh' % work_dir)
# misc.copy_file('data/src/speech/kaldi-run-chain-wrapper.sh',
# '%s/run-chain-wrapper.sh' % work_dir)
# misc.copy_file('data/src/speech/kaldi-run-chain-cfg.sh',
# '%s/run-chain-cfg.sh' % work_dir)
# misc.copy_file('data/src/speech/kaldi-run-chain-cpu.sh',
# '%s/run-chain-cpu.sh' % work_dir)
# misc.copy_file('data/src/speech/kaldi-run-chain-cpu-wrapper.sh',
# '%s/run-chain-cpu-wrapper.sh' % work_dir)
# misc.copy_file('data/src/speech/kaldi-run-chain-gpu.sh',
# '%s/run-chain-gpu.sh' % work_dir)
# misc.copy_file('data/src/speech/kaldi-run-chain-gpu-wrapper.sh',
# '%s/run-chain-gpu-wrapper.sh' % work_dir)
misc.copy_file('data/src/speech/kaldi-cmd.sh', '%s/cmd.sh' % work_dir)
misc.render_template('data/src/speech/kaldi-path.sh.template',
'%s/path.sh' % work_dir, kaldi_root=kaldi_root)
misc.mkdirs('%s/conf' % work_dir)
misc.copy_file('data/src/speech/kaldi-mfcc.conf',
'%s/conf/mfcc.conf' % work_dir)
misc.copy_file('data/src/speech/kaldi-mfcc-hires.conf',
'%s/conf/mfcc_hires.conf' % work_dir)
misc.copy_file('data/src/speech/kaldi-online-cmvn.conf',
'%s/conf/online_cmvn.conf' % work_dir)
misc.mkdirs('%s/local' % work_dir)
misc.copy_file('data/src/speech/kaldi-score.sh',
'%s/local/score.sh' % work_dir)
misc.mkdirs('%s/local/nnet3' % work_dir)
misc.copy_file('data/src/speech/kaldi-run-ivector-common.sh',
'%s/local/nnet3/run_ivector_common.sh' % work_dir)
def test_render_template(self):
# given
template_text = """VAR1={{val1}}
VAR2={{val2}}
"""
val1 = "v1"
val2 = "v2"
expected_text = """VAR1=%s
VAR2=%s
""" % (val1, val2)
src_path = os.path.join(str(self.test_dir), "src.txt")
dst_path = os.path.join(str(self.test_dir), "dst.txt")
with open(src_path, "wt") as f:
f.write(template_text)
# when
misc.render_template(src_path, dst_path, val1=val1, val2=val2)
# then
with open(dst_path) as f:
actual_text = f.read()
self.assertEqual(expected_text, actual_text)
cmd = 'rm -rf %s' % WORK_DIR
logging.debug(cmd)
os.system(cmd)
misc.mkdirs('%s/test' % data_dir)
#
# scripts
#
misc.render_template('data/src/speech/w2l_run_auto_review.sh.template',
'%s/run_auto_review.sh' % WORK_DIR,
w2l_env_activate=w2l_env_activate,
w2l_decoder=w2l_decoder,
cuda_device=CUDA_DEVICE,
w2l_tokensdir='../../data/models/%s' % model_name,
w2l_tokens='tokens.txt',
w2l_lexicon='../../data/models/%s/lexicon.txt' %
model_name,
w2l_am='../../data/models/%s/model.bin' % model_name,
w2l_lm='data/lm6.bin')
#
# read lexicon
#
lexfn = 'data/models/%s/lexicon.txt' % model_name
logging.info('reading lexicon %s ...' % lexfn)
lex = {}
def export_sphinx_case(work_dir, sphinxtrain_cfg_fn):
#
# language model
#
misc.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)
logging.info("%s written." % fn)
fn = '%s/wlist.txt' % work_dir
with codecs.open(fn, 'w', 'utf8') as outf:
for word in lex:
if ENABLE_NOISE_FILLER:
if word == NOISE_WORD:
logging.debug('skipping noise word')
continue
outf.write('%s\n' % word)
logging.info("%s written." % fn)
#
# create work_dir structure
#
mfcc_dir = "%s/mfcc" % work_dir
misc.mkdirs('%s/logs' % work_dir)
misc.mkdirs('%s/etc' % work_dir)
misc.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()
misc.copy_file(sphinxtrain_cfg_fn, '%s/etc/sphinx_train.cfg' % work_dir)
if ENABLE_NOISE_FILLER:
misc.copy_file('data/src/speech/sphinx-voxforge-noise.filler',
'%s/etc/voxforge.filler' % work_dir)
else:
misc.copy_file('data/src/speech/sphinx-voxforge.filler',
'%s/etc/voxforge.filler' % work_dir)
misc.copy_file('data/src/speech/sphinx-feat.params',
'%s/etc/feat.params' % work_dir)
#
# 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
lex_covered = set()
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/%s.wav" % (wav16_dir, cfn2corpus[cfn], cfn)
mfcfilename = "mfcc/%s.mfc" % cfn
runfeatf.write(SPHINXFE % (w16filename, mfcfilename, cnt))
cnt = (cnt + 1) % NJOBS
if cnt == 0:
runfeatf.write('wait\n')
logging.info("%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)
tokens = tokenize(ts_train[cfn]['ts'],
lang=options.lang,
keep_punctuation=False)
ts = u' '.join(tokens)
train_tsf.write(u'<s> %s </s> (%s)\n' % (ts, cfn))
for token in tokens:
if not token in lex:
logging.error('word %s not covered by dict!')
sys.exit(1)
lex_covered.add(token)
for cfn in ts_test:
test_fif.write('%s\n' % cfn)
tokens = tokenize(ts_test[cfn]['ts'],
lang=options.lang,
keep_punctuation=False)
ts = u' '.join(tokens)
test_tsf.write(u'<s> %s </s> (%s)\n' % (ts, cfn))
for token in tokens:
if not token in lex:
logging.error('word %s not covered by dict!')
sys.exit(1)
lex_covered.add(token)
logging.info("%s written." % train_tsfn)
logging.info("%s written." % train_fifn)
logging.info("%s written." % test_tsfn)
logging.info("%s written." % test_fifn)
# generate dict
phoneset = set()
pdfn = '%s/etc/voxforge.dic' % work_dir
with codecs.open(pdfn, 'w', 'utf8') as pdf:
for word in lex:
if ENABLE_NOISE_FILLER:
if word == NOISE_WORD:
logging.debug('skipping noise word')
continue
if not word in lex_covered:
logging.debug(
'skipping word %s as it is not covered by transcripts' %
word)
continue
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:
logging.error(
u"***ERROR: empty phone detected in lex entry %s %s" %
(word, ipa))
phoneset.add(phone)
logging.info("%s written." % pdfn)
logging.info("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')
if ENABLE_NOISE_FILLER:
phf.write(u'NSPC\n')
logging.info("%s written." % phfn)
misc.render_template('data/src/speech/sphinx-run.sh.template',
'%s/sphinx-run.sh' % work_dir,
lm_name=lm_name)
# clean up leftovers from previous runs
#
cmd = 'rm -rf %s' % WORKDIR
logging.info(cmd)
os.system(cmd)
misc.mkdirs(WORKDIR)
#
# copy scripts
#
misc.copy_file ('data/src/speech/kaldi-run-segmentation.sh', '%s/run-segmentation.sh' % WORKDIR)
misc.copy_file ('data/src/speech/kaldi-cmd.sh', '%s/cmd.sh' % WORKDIR)
misc.render_template ('data/src/speech/kaldi-path.sh.template', '%s/path.sh' % WORKDIR, kaldi_root=kaldi_root)
misc.symlink ('%s/egs/wsj/s5/steps' % kaldi_root, '%s/steps' % WORKDIR)
misc.symlink ('%s/egs/wsj/s5/utils' % kaldi_root, '%s/utils' % WORKDIR)
#
# create skeleton dst model
#
misc.mkdirs ('%s/exp' % WORKDIR)
cmd = "cp -r '%s/model' '%s/exp/tri2b_adapt'" % (modelfn, WORKDIR)
logging.info(cmd)
os.system(cmd)
cmd = "cp -r '%s/data' '%s/data'" % (modelfn, WORKDIR)
logging.info(cmd)
os.system(cmd)
def kaldi_adapt_lm(kaldi_root, src_model_dir, lm_fn, work_dir, dst_model_name):
steps_path = '%s/egs/wsj/s5/steps' % kaldi_root
if not os.path.exists (steps_path):
raise Exception ('%s does not exist - is kaldi really installed in %s ?' % (steps_path, kaldi_root))
tmpl_dir = os.path.dirname(os.path.abspath(__file__)) + '/templates'
#
# copy dictionary and phoneme sets from original model
#
logging.info("copying dictionary and phoneme sets from original model...")
misc.mkdirs('%s/data/local/dict' % work_dir)
misc.copy_file ('%s/data/local/dict/lexicon.txt' % src_model_dir, '%s/data/local/dict/lexicon.txt' % work_dir)
misc.copy_file ('%s/data/local/dict/nonsilence_phones.txt' % src_model_dir, '%s/data/local/dict/nonsilence_phones.txt' % work_dir)
misc.copy_file ('%s/data/local/dict/silence_phones.txt' % src_model_dir, '%s/data/local/dict/silence_phones.txt' % work_dir)
misc.copy_file ('%s/data/local/dict/optional_silence.txt' % src_model_dir, '%s/data/local/dict/optional_silence.txt' % work_dir)
misc.copy_file ('%s/data/local/dict/extra_questions.txt' % src_model_dir, '%s/data/local/dict/extra_questions.txt' % work_dir)
#
# language model
#
misc.copy_file (lm_fn, '%s/lm.arpa' % work_dir)
#
# create skeleton dst model
#
logging.info("creating skeleton destination model...")
misc.mkdirs ('%s/exp/adapt' % work_dir)
misc.copy_file ('%s/model/final.mdl' % src_model_dir, '%s/exp/adapt/final.mdl' % work_dir)
misc.copy_file ('%s/model/cmvn_opts' % src_model_dir, '%s/exp/adapt/cmvn_opts' % work_dir)
misc.copy_file ('%s/model/tree' % src_model_dir, '%s/exp/adapt/tree' % work_dir)
for optional_file in [ 'final.mat', 'splice_opts', 'final.occs', 'full.mat' ] :
if os.path.exists('%s/model/%s' % (src_model_dir, optional_file)):
misc.copy_file ('%s/model/%s' % (src_model_dir, optional_file), '%s/exp/adapt/%s' % (work_dir, optional_file))
if os.path.exists('%s/extractor' % src_model_dir):
misc.mkdirs ('%s/exp/extractor' % work_dir)
misc.copy_file ('%s/extractor/final.mat' % src_model_dir, '%s/exp/extractor/final.mat' % work_dir)
misc.copy_file ('%s/extractor/global_cmvn.stats' % src_model_dir, '%s/exp/extractor/global_cmvn.stats' % work_dir)
misc.copy_file ('%s/extractor/final.dubm' % src_model_dir, '%s/exp/extractor/final.dubm' % work_dir)
misc.copy_file ('%s/extractor/final.ie' % src_model_dir, '%s/exp/extractor/final.ie' % work_dir)
misc.copy_file ('%s/extractor/splice_opts' % src_model_dir, '%s/exp/extractor/splice_opts' % work_dir)
misc.mkdirs ('%s/exp/ivectors_test_hires/conf' % work_dir)
misc.copy_file ('%s/ivectors_test_hires/conf/splice.conf' % src_model_dir, '%s/exp/ivectors_test_hires/conf' % work_dir)
misc.mkdirs ('%s/conf' % work_dir)
misc.copy_file ('%s/conf/mfcc.conf' % src_model_dir, '%s/conf/mfcc.conf' % work_dir)
misc.copy_file ('%s/conf/mfcc_hires.conf' % src_model_dir, '%s/conf/mfcc_hires.conf' % work_dir)
misc.copy_file ('%s/conf/online_cmvn.conf' % src_model_dir, '%s/conf/online_cmvn.conf' % work_dir)
#
# copy scripts and config files
#
misc.copy_file ('%s/kaldi-run-adaptation.sh' % tmpl_dir, '%s/run-adaptation.sh' % work_dir)
misc.copy_file ('%s/kaldi-cmd.sh' % tmpl_dir, '%s/cmd.sh' % work_dir)
misc.render_template ('%s/kaldi-path.sh.template' % tmpl_dir, '%s/path.sh' % work_dir, kaldi_root=kaldi_root)
misc.copy_file ('%s/kaldi-model-dist.sh' % tmpl_dir, '%s/model-dist.sh' % work_dir)
misc.symlink ('%s/egs/wsj/s5/steps' % kaldi_root, '%s/steps' % work_dir)
misc.symlink ('%s/egs/wsj/s5/utils' % kaldi_root, '%s/utils' % work_dir)
cmd = '/bin/bash -c "pushd %s && bash run-adaptation.sh && popd"' % work_dir
logging.info (cmd)
os.system (cmd)
cmd = '/bin/bash -c "pushd %s && bash model-dist.sh "%s" && popd"' % (work_dir, dst_model_name)
logging.info (cmd)
os.system (cmd)
logging.info("loading lexicon...")
lex = Lexicon(file_name=dictionary)
logging.info("loading lexicon...done.")
#
# language model
#
misc.copy_file('%s/lm.arpa' % language_model_dir, '%s/lm.arpa' % data_dir)
#
# scripts
#
misc.render_template('data/src/speech/w2l_run_train.sh.template', '%s/run_train.sh' % work_dir, w2l_env_activate=w2l_env_activate, w2l_train=w2l_train)
misc.mkdirs('%s/config/conv_glu' % work_dir)
misc.render_template('data/src/speech/w2l_config_conv_glu_train.cfg.template', '%s/config/conv_glu/train.cfg' % work_dir, runname=model_name)
misc.copy_file('data/src/speech/w2l_config_conv_glu_network.arch', '%s/config/conv_glu/network.arch' % work_dir)
#
# export audio
#
def export_audio (train_val, tsdict):
global data_dir, utt_num, options
destdirfn = '%s/%s' % (data_dir, train_val)
for utt_id in tsdict: