Browse Source

label alignment using HVite is added.

master
yemaozi88 3 years ago
parent
commit
c185072d5b
  1. BIN
      .vs/acoustic_model/v15/.suo
  2. BIN
      acoustic_model/__pycache__/defaultfiles.cpython-36.pyc
  3. 2
      acoustic_model/acoustic_model.pyproj
  4. 6
      acoustic_model/convert_phoneset.py
  5. 1
      acoustic_model/defaultfiles.py
  6. 18
      acoustic_model/fame_functions.py
  7. 38
      acoustic_model/fame_hmm.py
  8. 815
      acoustic_model/htk_vs_kaldi.py
  9. 14
      acoustic_model/phoneset/fame_asr.py
  10. 22
      acoustic_model/stimmen_functions.py
  11. 18
      acoustic_model/stimmen_test.py

BIN
.vs/acoustic_model/v15/.suo

BIN
acoustic_model/__pycache__/defaultfiles.cpython-36.pyc

2
acoustic_model/acoustic_model.pyproj

@ -4,7 +4,7 @@
<SchemaVersion>2.0</SchemaVersion>
<ProjectGuid>4d8c8573-32f0-4a62-9e62-3ce5cc680390</ProjectGuid>
<ProjectHome>.</ProjectHome>
<StartupFile>fame_hmm.py</StartupFile>
<StartupFile>htk_vs_kaldi.py</StartupFile>
<SearchPath>
</SearchPath>
<WorkingDirectory>.</WorkingDirectory>

6
acoustic_model/convert_phoneset.py

@ -38,3 +38,9 @@ def convert_phoneset(word_list, translation_key):
translation_key (dict):
"""
return [translation_key.get(phone, phone) for phone in word_list]
def phone_reduction(phones, reduction_key):
multi_character_tokenize(wo.strip(), multi_character_phones)
return [reduction_key.get(i, i) for i in phones
if not i in phones_to_be_removed]

1
acoustic_model/defaultfiles.py

@ -17,6 +17,7 @@ novo_api_dir = os.path.join(WSL_dir, 'python-novo-api', 'novoapi')
rug_dir = r'c:\OneDrive\Research\rug'
experiments_dir = os.path.join(rug_dir, 'experiments')
htk_dir = os.path.join(experiments_dir, 'acoustic_model', 'fame', 'htk')
kaldi_dir = os.path.join(WSL_dir, 'kaldi-trunk', 'egs', '_stimmen')
stimmen_dir = os.path.join(experiments_dir, 'stimmen')
# data

18
acoustic_model/fame_functions.py

@ -321,9 +321,11 @@ def combine_lexicon(lexicon_file1, lexicon_file2, lexicon_out):
lex.to_csv(lexicon_out, index=False, header=False, sep='\t', encoding='utf-8')
def fix_single_quote(lexicon_file):
""" add '\' before all single quote at the beginning of words.
convert special characters to ascii compatible characters.
def fix_lexicon(lexicon_file):
""" fix lexicon
- add '\' before all single quote at the beginning of words.
- convert special characters to ascii compatible characters.
- add silence.
Args:
lexicon_file (path): lexicon file, which will be overwitten.
@ -331,6 +333,12 @@ def fix_single_quote(lexicon_file):
"""
lex = load_lexicon(lexicon_file)
lex = lex.dropna() # remove N/A.
# add 'sil'
row = pd.Series(['SILENCE', 'sil'], index=lex.columns)
lex = lex.append(row, ignore_index=True)
lex = lex.sort_values(by='word', ascending=True)
for i in lex[lex['word'].str.startswith('\'')].index.values:
lex.iat[i, 0] = lex.iat[i, 0].replace('\'', '\\\'')
# to_csv does not work with space seperator. therefore all tabs should manually be replaced.
@ -346,10 +354,11 @@ def word2htk(word):
def ipa2asr(ipa):
curr_dir = os.path.dirname(os.path.abspath(__file__))
translation_key_ipa2asr = np.load(os.path.join(curr_dir, 'phoneset', 'fame_ipa2asr.npy')).item(0)
#ipa_ = fame_asr.phone_reduction(ipa)
ipa_splitted = convert_phoneset.split_word(ipa, fame_ipa.multi_character_phones)
ipa_splitted = fame_ipa.phone_reduction(ipa_splitted)
asr_splitted = convert_phoneset.convert_phoneset(ipa_splitted, translation_key_ipa2asr)
asr_splitted = fame_asr.phone_reduction(asr_splitted)
return ''.join(asr_splitted)
@ -360,5 +369,6 @@ def ipa2htk(ipa):
ipa_splitted = convert_phoneset.split_word(ipa, fame_ipa.multi_character_phones)
ipa_splitted = fame_ipa.phone_reduction(ipa_splitted)
asr_splitted = convert_phoneset.convert_phoneset(ipa_splitted, translation_key_ipa2asr)
asr_splitted = fame_asr.phone_reduction(asr_splitted)
htk_splitted = convert_phoneset.convert_phoneset(asr_splitted, fame_asr.translation_key_asr2htk)
return ''.join(htk_splitted)

38
acoustic_model/fame_hmm.py

@ -27,7 +27,8 @@ extract_features = 0
flat_start = 0
train_model_without_sp = 0
add_sp = 0
train_model_with_sp = 1
train_model_with_sp = 0
train_model_with_sp_align_mlf = 1
@ -75,6 +76,7 @@ if not os.path.exists(label_dir):
## training
hcompv_scp_train = os.path.join(tmp_dir, 'train.scp')
mlf_file_train = os.path.join(label_dir, 'train_phone.mlf')
mlf_file_train_aligned = os.path.join(label_dir, 'train_phone_aligned.mlf')
## train without sp
niter_max = 10
@ -102,7 +104,8 @@ if make_lexicon:
# (1) Replace all tabs with single space;
# (2) Put a '\' before any dictionary entry beginning with single quote
#http://electroblaze.blogspot.nl/2013/03/understanding-htk-error-messages.html
fame_functions.fix_single_quote(lexicon_htk)
print('>>> fixing the lexicon...')
fame_functions.fix_lexicon(lexicon_htk)
print("elapsed time: {}".format(time.time() - timer_start))
@ -269,11 +272,11 @@ if train_model_without_sp:
fh.make_new_directory(modeln_dir)
pyhtk.re_estimation(
config_train,
os.path.join(modeln_dir_pre, 'macros'),
os.path.join(modeln_dir_pre, hmmdefs_name),
modeln_dir,
hcompv_scp_train, phonelist_txt,
mlf_file=mlf_file_train)
mlf_file=mlf_file_train,
macros=os.path.join(modeln_dir_pre, 'macros'))
print("elapsed time: {}".format(time.time() - timer_start))
@ -321,7 +324,6 @@ if add_sp:
## ======================= train model with short pause =======================
if train_model_with_sp:
print('==== train model with sp ====')
#for niter in range(niter_max+1, niter_max*2+1):
for niter in range(20, 50):
timer_start = time.time()
hmm_n = 'iter' + str(niter)
@ -333,9 +335,31 @@ if train_model_with_sp:
fh.make_new_directory(modeln_dir)
pyhtk.re_estimation(
config_train,
os.path.join(modeln_dir_pre, 'macros'),
os.path.join(modeln_dir_pre, hmmdefs_name),
modeln_dir,
hcompv_scp_train, phonelist_txt,
mlf_file=mlf_file_train)
mlf_file=mlf_file_train,
macros=os.path.join(modeln_dir_pre, 'macros'))
print("elapsed time: {}".format(time.time() - timer_start))
## ======================= train model with short pause =======================
if train_model_with_sp_align_mlf:
print('==== train model with sp with align.mlf ====')
for niter in range(50, 60):
timer_start = time.time()
hmm_n = 'iter' + str(niter)
hmm_n_pre = 'iter' + str(niter-1)
modeln_dir = os.path.join(model1_dir, hmm_n)
modeln_dir_pre = os.path.join(model1_dir, hmm_n_pre)
# re-estimation
fh.make_new_directory(modeln_dir)
pyhtk.re_estimation(
config_train,
os.path.join(modeln_dir_pre, hmmdefs_name),
modeln_dir,
hcompv_scp_train, phonelist_txt,
mlf_file=mlf_file_train_aligned,
macros=os.path.join(modeln_dir_pre, 'macros'))
print("elapsed time: {}".format(time.time() - timer_start))

815
acoustic_model/htk_vs_kaldi.py

@ -11,6 +11,7 @@ import glob
import numpy as np
import pandas as pd
from collections import Counter
#import matplotlib.pyplot as plt
#from sklearn.metrics import confusion_matrix
@ -50,11 +51,14 @@ from htk import pyhtk
#lex_asr = os.path.join(default.fame_dir, 'lexicon', 'lex.asr')
#lex_asr_htk = os.path.join(default.fame_dir, 'lexicon', 'lex.asr_htk')
## procedure
# procedure
make_dic_file = 0
make_HTK_files = 1
extract_features = 0
#make_htk_dict_files = 0
#do_forced_alignment_htk = 0
#eval_forced_alignment_htk = 0
#make_kaldi_data_files = 0
make_kaldi_files = 0
#make_kaldi_lexicon_txt = 0
#load_forced_alignment_kaldi = 1
#eval_forced_alignment_kaldi = 1
@ -66,13 +70,34 @@ from htk import pyhtk
#sys.path.append(os.path.join(default.repo_dir, 'toolbox'))
#from evaluation import plot_confusion_matrix
config_dir = os.path.join(default.htk_dir, 'config')
model_dir = os.path.join(default.htk_dir, 'model')
lattice_file = os.path.join(config_dir, 'stimmen.ltc')
#pyhtk.create_word_lattice_file(
# os.path.join(config_dir, 'stimmen.net'),
# lattice_file)
hvite_scp = os.path.join(default.htk_dir, 'tmp', 'stimmen_test.scp')
## HTK related files.
config_dir = os.path.join(default.htk_dir, 'config')
model_dir = os.path.join(default.htk_dir, 'model')
feature_dir = os.path.join(default.htk_dir, 'mfc', 'stimmen')
config_hcopy = os.path.join(config_dir, 'config.HCopy')
# files to be made.
lattice_file = os.path.join(config_dir, 'stimmen.ltc')
phonelist_txt = os.path.join(config_dir, 'phonelist.txt')
stimmen_dic = os.path.join(default.htk_dir, 'lexicon', 'stimmen_recognition.dic')
hcopy_scp = os.path.join(default.htk_dir, 'tmp', 'stimmen_test_hcopy.scp')
hvite_scp = os.path.join(default.htk_dir, 'tmp', 'stimmen_test_hvite.scp')
hresult_scp = os.path.join(default.htk_dir, 'tmp', 'stimmen_test_result.scp')
## Kaldi related files.
kaldi_data_dir = os.path.join(default.kaldi_dir, 'data')
# files to be made.
wav_scp = os.path.join(kaldi_data_dir, 'test', 'wav.scp')
text_file = os.path.join(kaldi_data_dir, 'test', 'text')
utt2spk = os.path.join(kaldi_data_dir, 'test', 'utt2spk')
corpus_txt = os.path.join(kaldi_data_dir, 'local', 'corpus.txt')
lexicon_txt = os.path.join(kaldi_data_dir, 'local', 'dict', 'lexicon.txt')
nonsilence_phones_txt = os.path.join(kaldi_data_dir, 'local', 'dict', 'nonsilence_phones.txt')
silence_phones_txt = os.path.join(kaldi_data_dir, 'local', 'dict', 'silence_phones.txt')
optional_silence_txt = os.path.join(kaldi_data_dir, 'local', 'dict', 'optional_silence.txt')
## ======================= load test data ======================
@ -85,392 +110,468 @@ df = stimmen_functions.add_row_htk(df)
word_list = [i for i in list(set(df['word'])) if not pd.isnull(i)]
word_list = sorted(word_list)
# pronunciation variants
## ======================= make dic file to check pronunciation variants ======================
# dic file should be manually modified depends on the task - recognition / forced-alignemnt.
if make_dic_file:
# for HTK.
with open(stimmen_dic, mode='wb') as f:
for word in word_list:
df_ = df[df['word']==word]
pronunciations = list(np.unique(df_['htk']))
pronunciations_ = [word.upper() + ' sil ' + ' '.join(convert_phoneset.split_word(
htk, fame_asr.multi_character_phones_htk)) + ' sil'
for htk in pronunciations]
f.write(bytes('\n'.join(pronunciations_) + '\n', 'ascii'))
f.write(bytes('SILENCE sil\n', 'ascii'))
# for Kaldi.
fh.make_new_directory(os.path.join(kaldi_data_dir, 'local', 'dict'))
with open(lexicon_txt, mode='wb') as f:
f.write(bytes('!SIL sil\n', 'utf-8'))
f.write(bytes('<UNK> spn\n', 'utf-8'))
for word in word_list:
df_ = df[df['word']==word]
pronunciations = list(np.unique(df_['asr']))
pronunciations_ = [word.lower() + ' ' + ' '.join(convert_phoneset.split_word(
asr, fame_asr.multi_character_phones))
for asr in pronunciations]
f.write(bytes('\n'.join(pronunciations_) + '\n', 'utf-8'))
## ======================= test data for recognition ======================
# only target pronunciation variants.
df_rec = pd.DataFrame(index=[], columns=list(df.keys()))
for word in word_list:
df_ = df[df['word']==word]
print('{0} has {1} variants'.format(word, len(np.unique(df_['htk'])))
#fh.make_filelist(stimmen_test_dir, hvite_scp, file_type='wav')
#output = pyhtk.recognition(
# os.path.join(default.htk_dir, 'config', 'config.rec',
# lattice_file,
# os.path.join(model_dir, 'hmm1', 'iter13'),
# dictionary_file,
# os.path.join(config_dir, 'phonelist.txt'),
# hvite_scp)
#pyhtk.create_label_file(
# row['word'],
# os.path.join(stimmen_test_dir, filename.replace('.wav', '.lab')))
## ======================= make a HTK dic file ======================
#if make_htk_dic_file:
# output_type = 3
dictionary_txt = os.path.join(default.htk_dir, 'lexicon', 'stimmen.dic')
#for word in word_list:
word = word_list[2]
# pronunciation variant of the target word.
pronunciations = df_test['asr'][df_test['word'].str.match(word)]
# make dic file.
#am_func.make_htk_dict(word, pronvar_, htk_dict_file, output_type)
variants = [htk.replace(' ', '')
for htk in stimmen_functions.load_pronunciations(word.upper(), stimmen_dic)]
df_ = df[df['word'] == word]
for index, row in df_.iterrows():
if row['htk'] in variants:
df_rec = df_rec.append(row, ignore_index=True)
## ======================= make files required for HTK ======================
if make_HTK_files:
# make a word lattice file.
pyhtk.create_word_lattice_file(
os.path.join(config_dir, 'stimmen.net'),
lattice_file)
# extract features.
with open(hcopy_scp, 'wb') as f:
filelist = [os.path.join(stimmen_test_dir, filename) + '\t'
+ os.path.join(feature_dir, os.path.basename(filename).replace('.wav', '.mfc'))
for filename in df['filename']]
f.write(bytes('\n'.join(filelist), 'ascii'))
pyhtk.wav2mfc(config_hcopy, hcopy_scp)
# make label files.
for index, row in df.iterrows():
filename = row['filename'].replace('.wav', '.lab')
label_file = os.path.join(feature_dir, filename)
with open(label_file, 'wb') as f:
label_string = 'START\n' + row['word'].upper() + '\nEND\n'
f.write(bytes(label_string, 'ascii'))
## ======================= make files required for Kaldi =======================
if make_kaldi_files:
fh.make_new_directory(os.path.join(kaldi_data_dir, 'test'))
fh.make_new_directory(os.path.join(kaldi_data_dir, 'test', 'local'))
fh.make_new_directory(os.path.join(kaldi_data_dir, 'conf'))
# remove previous files.
if os.path.exists(wav_scp):
os.remove(wav_scp)
if os.path.exists(text_file):
os.remove(text_file)
if os.path.exists(utt2spk):
os.remove(utt2spk)
f_wav_scp = open(wav_scp, 'a', encoding="utf-8", newline='\n')
f_text_file = open(text_file, 'a', encoding="utf-8", newline='\n')
f_utt2spk = open(utt2spk, 'a', encoding="utf-8", newline='\n')
# make wav.scp, text, and utt2spk files.
for i, row in df_rec.iterrows():
filename = row['filename']
print('=== {0}: {1} ==='.format(i, filename))
wav_file = os.path.join(stimmen_test_dir, filename)
#if os.path.exists(wav_file):
speaker_id = 'speaker_' + str(i).zfill(4)
utterance_id = filename.replace('.wav', '')
utterance_id = utterance_id.replace(' ', '_')
utterance_id = speaker_id + '-' + utterance_id
# output
f_wav_scp.write('{0} {1}\n'.format(
utterance_id,
wav_file.replace('c:/', '/mnt/c/').replace('\\', '/'))) # convert path to unix format.
f_text_file.write('{0}\t{1}\n'.format(utterance_id, df_rec['word'][i].lower()))
f_utt2spk.write('{0} {1}\n'.format(utterance_id, speaker_id))
f_wav_scp.close()
f_text_file.close()
f_utt2spk.close()
with open(corpus_txt, 'wb') as f:
f.write(bytes('\n'.join([word.lower() for word in word_list]) + '\n', 'utf-8'))
with open(nonsilence_phones_txt, 'wb') as f:
f.write(bytes('\n'.join(fame_asr.phoneset_short) + '\n', 'utf-8'))
with open(silence_phones_txt, 'wb') as f:
f.write(bytes('sil\nspn\n', 'utf-8'))
with open(optional_silence_txt, 'wb') as f:
f.write(bytes('sil\n', 'utf-8'))
with open(os.path.join(kaldi_data_dir, 'conf', 'decode.config'), 'wb') as f:
f.write(bytes('first_beam=10.0\n', 'utf-8'))
f.write(bytes('beam=13.0\n', 'utf-8'))
f.write(bytes('lattice_beam=6.0\n', 'utf-8'))
with open(os.path.join(kaldi_data_dir, 'conf', 'mfcc.conf'), 'wb') as f:
f.write(bytes('--use-energy=false', 'utf-8'))
## ======================= recognition ======================
listdir = glob.glob(os.path.join(feature_dir, '*.mfc'))
with open(hvite_scp, 'wb') as f:
f.write(bytes('\n'.join(listdir), 'ascii'))
with open(hresult_scp, 'wb') as f:
f.write(bytes('\n'.join(listdir).replace('.mfc', '.rec'), 'ascii'))
# calculate result
performance = np.zeros((1, 2))
for niter in range(1, 50):
output = pyhtk.recognition(
os.path.join(config_dir, 'config.rec'),
lattice_file,
os.path.join(default.htk_dir, 'model', 'hmm1', 'iter' + str(niter), 'hmmdefs'),
stimmen_dic, phonelist_txt, hvite_scp)
output = pyhtk.calc_recognition_performance(
stimmen_dic, hresult_scp)
per_sentence, per_word = pyhtk.load_recognition_output_all(output)
performance_ = np.array([niter, per_sentence['accuracy']]).reshape(1, 2)
performance = np.r_[performance, performance_]
print('{0}: {1}[%]'.format(niter, per_sentence['accuracy']))
## ======================= forced alignment using HTK =======================
if do_forced_alignment_htk:
#for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]:
for hmm_num in [256, 512, 1024]:
hmm_num_str = str(hmm_num)
acoustic_model = os.path.join(acoustic_model_dir, 'hmm' + hmm_num_str + r'-2\hmmdefs')
predictions = pd.DataFrame({'filename': [''],
'word': [''],
'xsampa': [''],
'ipa': [''],
'famehtk': [''],
'prediction': ['']})
for i, filename in enumerate(df['filename']):
print('=== {0}/{1} ==='.format(i, len(df)))
if (i in df['filename'].keys()) and (isinstance(df['filename'][i], str)):
wav_file = os.path.join(wav_dir, filename)
if os.path.exists(wav_file):
word = df['word'][i]
WORD = word.upper()
fa_file = os.path.join(fa_dir, filename.replace('.wav', '.txt') + hmm_num_str)
#if not os.path.exists(fa_file):
# make label file.
label_file = os.path.join(wav_dir, filename.replace('.wav', '.lab'))
with open(label_file, 'w') as f:
lines = f.write(WORD)
htk_dict_file = os.path.join(htk_dict_dir, word + '.dic')
pyhtk.doHVite(wav_file, label_file, htk_dict_file, fa_file, default.config_hvite,
default.phonelist, acoustic_model)
os.remove(label_file)
prediction = am_func.read_fileFA(fa_file)
print('{0}: {1} -> {2}'.format(WORD, df['famehtk'][i], prediction))
else:
prediction = ''
print('!!!!! file not found.')
line = pd.Series([df['filename'][i], df['word'][i], df['xsampa'][i], df['ipa'][i], df['famehtk'][i], prediction], index=['filename', 'word', 'xsampa', 'ipa', 'famehtk', 'prediction'], name=i)
predictions = predictions.append(line)
else:
prediction = ''
print('!!!!! invalid entry.')
predictions.to_pickle(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.pkl'))
## ======================= make files which is used for forced alignment by Kaldi =======================
if make_kaldi_data_files:
wav_scp = os.path.join(kaldi_data_dir, 'wav.scp')
text_file = os.path.join(kaldi_data_dir, 'text')
utt2spk = os.path.join(kaldi_data_dir, 'utt2spk')
# remove previous files.
if os.path.exists(wav_scp):
os.remove(wav_scp)
if os.path.exists(text_file):
os.remove(text_file)
if os.path.exists(utt2spk):
os.remove(utt2spk)
f_wav_scp = open(wav_scp, 'a', encoding="utf-8", newline='\n')
f_text_file = open(text_file, 'a', encoding="utf-8", newline='\n')
f_utt2spk = open(utt2spk, 'a', encoding="utf-8", newline='\n')
# make wav.scp, text, and utt2spk files.
for i in df.index:
filename = df['filename'][i]
print('=== {0}: {1} ==='.format(i, filename))
#if (i in df['filename'].keys()) and (isinstance(df['filename'][i], str)):
wav_file = os.path.join(wav_dir, filename)
if os.path.exists(wav_file):
speaker_id = 'speaker_' + str(i).zfill(4)
utterance_id = filename.replace('.wav', '')
utterance_id = utterance_id.replace(' ', '_')
utterance_id = speaker_id + '-' + utterance_id
# wav.scp file
wav_file_unix = wav_file.replace('\\', '/')
wav_file_unix = wav_file_unix.replace('c:/', '/mnt/c/')
f_wav_scp.write('{0} {1}\n'.format(utterance_id, wav_file_unix))
# text file
word = df['word'][i].lower()
f_text_file.write('{0}\t{1}\n'.format(utterance_id, word))
# utt2spk
f_utt2spk.write('{0} {1}\n'.format(utterance_id, speaker_id))
f_wav_scp.close()
f_text_file.close()
f_utt2spk.close()
#for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]:
for hmm_num in [256, 512, 1024]:
hmm_num_str = str(hmm_num)
acoustic_model = os.path.join(acoustic_model_dir, 'hmm' + hmm_num_str + r'-2\hmmdefs')
predictions = pd.DataFrame({'filename': [''],
'word': [''],
'xsampa': [''],
'ipa': [''],
'famehtk': [''],
'prediction': ['']})
for i, filename in enumerate(df['filename']):
print('=== {0}/{1} ==='.format(i, len(df)))
if (i in df['filename'].keys()) and (isinstance(df['filename'][i], str)):
wav_file = os.path.join(wav_dir, filename)
if os.path.exists(wav_file):
word = df['word'][i]
WORD = word.upper()
fa_file = os.path.join(fa_dir, filename.replace('.wav', '.txt') + hmm_num_str)
#if not os.path.exists(fa_file):
# make label file.
label_file = os.path.join(wav_dir, filename.replace('.wav', '.lab'))
with open(label_file, 'w') as f:
lines = f.write(WORD)
htk_dict_file = os.path.join(htk_dict_dir, word + '.dic')
pyhtk.doHVite(wav_file, label_file, htk_dict_file, fa_file, default.config_hvite,
default.phonelist, acoustic_model)
os.remove(label_file)
prediction = am_func.read_fileFA(fa_file)
print('{0}: {1} -> {2}'.format(WORD, df['famehtk'][i], prediction))
else:
prediction = ''
print('!!!!! file not found.')
line = pd.Series([df['filename'][i], df['word'][i], df['xsampa'][i], df['ipa'][i], df['famehtk'][i], prediction], index=['filename', 'word', 'xsampa', 'ipa', 'famehtk', 'prediction'], name=i)
predictions = predictions.append(line)
else:
prediction = ''
print('!!!!! invalid entry.')
predictions.to_pickle(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.pkl'))
## ======================= make lexicon txt which is used by Kaldi =======================
if make_kaldi_lexicon_txt:
option_num = 6
option_num = 6
# remove previous file.
if os.path.exists(lexicon_txt):
os.remove(lexicon_txt)
lexiconp_txt = lexicon_txt.replace('lexicon.txt', 'lexiconp.txt')
if os.path.exists(lexiconp_txt):
os.remove(lexiconp_txt)
# output lexicon.txt
f_lexicon_txt = open(lexicon_txt, 'a', encoding="utf-8", newline='\n')
pronvar_list_all = []
for word in word_list:
# remove previous file.
if os.path.exists(lexicon_txt):
os.remove(lexicon_txt)
lexiconp_txt = lexicon_txt.replace('lexicon.txt', 'lexiconp.txt')
if os.path.exists(lexiconp_txt):
os.remove(lexiconp_txt)
# output lexicon.txt
f_lexicon_txt = open(lexicon_txt, 'a', encoding="utf-8", newline='\n')
pronvar_list_all = []
for word in word_list:
# pronunciation variant of the target word.
pronunciation_variants = df['ipa'][df['word'].str.match(word)]
# pronunciation variant of the target word.
pronunciation_variants = df['ipa'][df['word'].str.match(word)]
c = Counter(pronunciation_variants)
total_num = sum(c.values())
c = Counter(pronunciation_variants)
total_num = sum(c.values())
#with open(result_dir + '\\' + word + '.csv', 'a', encoding="utf-8", newline='\n') as f:
# for key in c.keys():
# f.write("{0},{1}\n".format(key,c[key]))
#with open(result_dir + '\\' + word + '.csv', 'a', encoding="utf-8", newline='\n') as f:
# for key in c.keys():
# f.write("{0},{1}\n".format(key,c[key]))
for key, value in c.most_common(option_num):
# make possible pronunciation variant list.
pronvar_list = am_func.fame_pronunciation_variant(key)
for key, value in c.most_common(option_num):
# make possible pronunciation variant list.
pronvar_list = am_func.fame_pronunciation_variant(key)
for pronvar_ in pronvar_list:
split_ipa = convert_phone_set.split_fame_ipa(pronvar_)
pronvar_out = ' '.join(split_ipa)
pronvar_list_all.append([word, pronvar_out])
for pronvar_ in pronvar_list:
split_ipa = convert_phone_set.split_fame_ipa(pronvar_)
pronvar_out = ' '.join(split_ipa)
pronvar_list_all.append([word, pronvar_out])
pronvar_list_all = np.array(pronvar_list_all)
pronvar_list_all = np.unique(pronvar_list_all, axis=0)
pronvar_list_all = np.array(pronvar_list_all)
pronvar_list_all = np.unique(pronvar_list_all, axis=0)
# output
f_lexicon_txt.write('<UNK>\tSPN\n')
for line in pronvar_list_all:
f_lexicon_txt.write('{0}\t{1}\n'.format(line[0].lower(), line[1]))
# output
f_lexicon_txt.write('<UNK>\tSPN\n')
for line in pronvar_list_all:
f_lexicon_txt.write('{0}\t{1}\n'.format(line[0].lower(), line[1]))
f_lexicon_txt.close()
f_lexicon_txt.close()
## ======================= load kaldi forced alignment result =======================
if load_forced_alignment_kaldi:
phones_txt = os.path.join(default.kaldi_dir, 'data', 'lang', 'phones.txt')
merged_alignment_txt = os.path.join(default.kaldi_dir, 'exp', 'tri1_alignme', 'merged_alignment.txt')
#filenames = np.load(data_dir + '\\filenames.npy')
#words = np.load(data_dir + '\\words.npy')
#pronunciations = np.load(data_dir + '\\pronunciations_ipa.npy')
#pronvar_list_all = np.load(data_dir + '\\pronvar_list_all.npy')
#word_list = np.unique(words)
# load the mapping between phones and ids.
with open(phones_txt, 'r', encoding="utf-8") as f:
mapping_phone2id = f.read().split('\n')
phones = []
phone_ids = [] # ID of phones
for m in mapping_phone2id:
m = m.split(' ')
if len(m) > 1:
phones.append(m[0])
phone_ids.append(int(m[1]))
# load the result of FA.
with open(merged_alignment_txt, 'r') as f:
lines = f.read()
lines = lines.split('\n')
predictions = pd.DataFrame({'filename': [''],
'word': [''],
'xsampa': [''],
'ipa': [''],
'famehtk': [''],
'prediction': ['']})
#fa_filenames = []
#fa_pronunciations = []
utterance_id_ = ''
pronunciation = []
for line in lines:
line = line.split(' ')
if len(line) == 5:
utterance_id = line[0]
if utterance_id == utterance_id_:
phone_id = int(line[4])
#if not phone_id == 1:
phone_ = phones[phone_ids.index(phone_id)]
phone = re.sub(r'_[A-Z]', '', phone_)
if not phone == 'SIL':
pronunciation.append(phone)
else:
filename = re.sub(r'speaker_[0-9]{4}-', '', utterance_id_)
prediction = ''.join(pronunciation)
df_ = df[df['filename'].str.match(filename)]
df_idx = df_.index[0]
prediction_ = pd.Series([#filename,
#df_['word'][df_idx],
#df_['xsampa'][df_idx],
#df_['ipa'][df_idx],
#df_['famehtk'][df_idx],
df_.iloc[0,1],
df_.iloc[0,3],
df_.iloc[0,4],
df_.iloc[0,2],
df_.iloc[0,0],
prediction],
index=['filename', 'word', 'xsampa', 'ipa', 'famehtk', 'prediction'],
name=df_idx)
predictions = predictions.append(prediction_)
#fa_filenames.append()
#fa_pronunciations.append(' '.join(pronunciation))
pronunciation = []
utterance_id_ = utterance_id
predictions.to_pickle(os.path.join(result_dir, 'kaldi', 'predictions.pkl'))
phones_txt = os.path.join(default.kaldi_dir, 'data', 'lang', 'phones.txt')
merged_alignment_txt = os.path.join(default.kaldi_dir, 'exp', 'tri1_alignme', 'merged_alignment.txt')
#filenames = np.load(data_dir + '\\filenames.npy')
#words = np.load(data_dir + '\\words.npy')
#pronunciations = np.load(data_dir + '\\pronunciations_ipa.npy')
#pronvar_list_all = np.load(data_dir + '\\pronvar_list_all.npy')
#word_list = np.unique(words)
# load the mapping between phones and ids.
with open(phones_txt, 'r', encoding="utf-8") as f:
mapping_phone2id = f.read().split('\n')
phones = []
phone_ids = [] # ID of phones
for m in mapping_phone2id:
m = m.split(' ')
if len(m) > 1:
phones.append(m[0])
phone_ids.append(int(m[1]))
# load the result of FA.
with open(merged_alignment_txt, 'r') as f:
lines = f.read()
lines = lines.split('\n')
predictions = pd.DataFrame({'filename': [''],
'word': [''],
'xsampa': [''],
'ipa': [''],
'famehtk': [''],
'prediction': ['']})
#fa_filenames = []
#fa_pronunciations = []
utterance_id_ = ''
pronunciation = []
for line in lines:
line = line.split(' ')
if len(line) == 5:
utterance_id = line[0]
if utterance_id == utterance_id_:
phone_id = int(line[4])
#if not phone_id == 1:
phone_ = phones[phone_ids.index(phone_id)]
phone = re.sub(r'_[A-Z]', '', phone_)
if not phone == 'SIL':
pronunciation.append(phone)
else:
filename = re.sub(r'speaker_[0-9]{4}-', '', utterance_id_)
prediction = ''.join(pronunciation)
df_ = df[df['filename'].str.match(filename)]
df_idx = df_.index[0]
prediction_ = pd.Series([#filename,
#df_['word'][df_idx],
#df_['xsampa'][df_idx],
#df_['ipa'][df_idx],
#df_['famehtk'][df_idx],
df_.iloc[0,1],
df_.iloc[0,3],
df_.iloc[0,4],
df_.iloc[0,2],
df_.iloc[0,0],
prediction],
index=['filename', 'word', 'xsampa', 'ipa', 'famehtk', 'prediction'],
name=df_idx)
predictions = predictions.append(prediction_)
#fa_filenames.append()
#fa_pronunciations.append(' '.join(pronunciation))
pronunciation = []
utterance_id_ = utterance_id
predictions.to_pickle(os.path.join(result_dir, 'kaldi', 'predictions.pkl'))
## ======================= evaluate the result of forced alignment =======================
if eval_forced_alignment_htk:
htk_dict_dir = os.path.join(default.experiments_dir, 'stimmen', 'dic_short')
htk_dict_dir = os.path.join(default.experiments_dir, 'stimmen', 'dic_short')
compare_hmm_num = 1
compare_hmm_num = 1
if compare_hmm_num:
f_result = open(os.path.join(result_dir, 'result.csv'), 'w')
f_result.write("nmix,Oog,Oog,Oor,Oor,Pauw,Pauw,Reus,Reus,Reuzenrad,Reuzenrad,Roeiboot,Roeiboot,Rozen,Rozen\n")
if compare_hmm_num:
f_result = open(os.path.join(result_dir, 'result.csv'), 'w')
f_result.write("nmix,Oog,Oog,Oor,Oor,Pauw,Pauw,Reus,Reus,Reuzenrad,Reuzenrad,Roeiboot,Roeiboot,Rozen,Rozen\n")
for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]:
#for hmm_num in [256]:
hmm_num_str = str(hmm_num)
if compare_hmm_num:
f_result.write("{},".format(hmm_num_str))
for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]:
#for hmm_num in [256]:
hmm_num_str = str(hmm_num)
if compare_hmm_num:
f_result.write("{},".format(hmm_num_str))
#match = np.load(data_dir + '\\match_hmm' + hmm_num_str + '.npy')
#prediction = np.load(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.npy'))
#prediction = pd.Series(prediction, index=df.index, name='prediction')
#result = pd.concat([df, prediction], axis=1)
result = pd.read_pickle(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.pkl'))
#match = np.load(data_dir + '\\match_hmm' + hmm_num_str + '.npy')
#prediction = np.load(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.npy'))
#prediction = pd.Series(prediction, index=df.index, name='prediction')
#result = pd.concat([df, prediction], axis=1)
result = pd.read_pickle(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.pkl'))
# load pronunciation variants
for word in word_list:
htk_dict_file = os.path.join(htk_dict_dir, word + '.dic')
with open(htk_dict_file, 'r') as f:
lines = f.read().split('\n')[:-1]
pronunciation_variants = [line.split('\t')[1] for line in lines]
# load pronunciation variants
for word in word_list:
htk_dict_file = os.path.join(htk_dict_dir, word + '.dic')
with open(htk_dict_file, 'r') as f:
lines = f.read().split('\n')[:-1]
pronunciation_variants = [line.split('\t')[1] for line in lines]
# see only words which appears in top 3.
result_ = result[result['word'].str.match(word)]
result_ = result_[result_['famehtk'].isin(pronunciation_variants)]
# see only words which appears in top 3.
result_ = result[result['word'].str.match(word)]
result_ = result_[result_['famehtk'].isin(pronunciation_variants)]
match_num = sum(result_['famehtk'] == result_['prediction'])
total_num = len(result_)
match_num = sum(result_['famehtk'] == result_['prediction'])
total_num = len(result_)
print("word '{0}': {1}/{2} ({3:.2f} %)".format(word, match_num, total_num, match_num/total_num*100))
if compare_hmm_num:
f_result.write("{0},{1},".format(match_num, total_num))
else:
# output confusion matrix
cm = confusion_matrix(result_['famehtk'], result_['prediction'])
print("word '{0}': {1}/{2} ({3:.2f} %)".format(word, match_num, total_num, match_num/total_num*100))
if compare_hmm_num:
f_result.write("{0},{1},".format(match_num, total_num))
else:
# output confusion matrix
cm = confusion_matrix(result_['famehtk'], result_['prediction'])
plt.figure()
plot_confusion_matrix(cm, classes=pronunciation_variants, normalize=False)
plt.savefig(result_dir + '\\cm_' + word + '.png')
plt.figure()
plot_confusion_matrix(cm, classes=pronunciation_variants, normalize=False)
plt.savefig(result_dir + '\\cm_' + word + '.png')
if compare_hmm_num:
f_result.write('\n')
if compare_hmm_num:
f_result.write('\n')
if compare_hmm_num:
f_result.close()
if compare_hmm_num:
f_result.close()
## ======================= evaluate the result of forced alignment of kaldi =======================
if eval_forced_alignment_kaldi:
result = pd.read_pickle(os.path.join(result_dir, 'kaldi', 'predictions.pkl'))
f_result = open(os.path.join(result_dir, 'result.csv'), 'w')
f_result.write("word,total,valid,match,[%]\n")
# load pronunciation variants
with open(lexicon_txt, 'r', encoding="utf-8", newline='\n') as f:
lines = f.read().split('\n')[:-1]
pronunciation_variants_all = [line.split('\t') for line in lines]
word_list = np.delete(word_list, [0], 0) # remove 'Oog'
for word in word_list:
# load pronunciation variant of the word.
pronunciation_variants = []
for line in pronunciation_variants_all:
if line[0] == word.lower():
pronunciation_variants.append(line[1].replace(' ', ''))
# see only words which appears in top 3.
result_ = result[result['word'].str.match(word)]
result_tolerant = pd.DataFrame({
'filename': [''],
'word': [''],
'xsampa': [''],
'ipa': [''],
'prediction': [''],
'match': ['']})
for i in range(0, len(result_)):
line = result_.iloc[i]
# make a list of all possible pronunciation variants of ipa description.
# i.e. possible answers from forced alignment.
ipa = line['ipa']
pronvar_list = [ipa]
pronvar_list_ = am_func.fame_pronunciation_variant(ipa)
if not pronvar_list_ is None:
pronvar_list += list(pronvar_list_)
# only focus on pronunciations which can be estimated from ipa.
if len(set(pronvar_list) & set(pronunciation_variants)) > 0:
if line['prediction'] in pronvar_list:
ismatch = True
else:
ismatch = False
line_df = pd.DataFrame(result_.iloc[i]).T
df_idx = line_df.index[0]
result_tolerant_ = pd.Series([line_df.loc[df_idx, 'filename'],
line_df.loc[df_idx, 'word'],
line_df.loc[df_idx, 'xsampa'],
line_df.loc[df_idx, 'ipa'],
line_df.loc[df_idx, 'prediction'],
ismatch],
index=['filename', 'word', 'xsampa', 'ipa', 'prediction', 'match'],
name=df_idx)
result_tolerant = result_tolerant.append(result_tolerant_)
# remove the first entry (dummy)
result_tolerant = result_tolerant.drop(0, axis=0)
total_num = len(result_)
valid_num = len(result_tolerant)
match_num = np.sum(result_tolerant['match'])
print("word '{0}': {1}/{2} ({3:.2f} %) originally {4}".format(word, match_num, valid_num, match_num/valid_num*100, total_num))
f_result.write("{0},{1},{2},{3},{4}\n".format(word, total_num, valid_num, match_num, match_num/valid_num*100))
f_result.close()
## output confusion matrix
#cm = confusion_matrix(result_['ipa'], result_['prediction'])
#plt.figure()
#plot_confusion_matrix(cm, classes=pronunciation_variants, normalize=False)
#plt.savefig(result_dir + '\\cm_' + word + '.png')
result = pd.read_pickle(os.path.join(result_dir, 'kaldi', 'predictions.pkl'))
f_result = open(os.path.join(result_dir, 'result.csv'), 'w')
f_result.write("word,total,valid,match,[%]\n")
# load pronunciation variants
with open(lexicon_txt, 'r', encoding="utf-8", newline='\n') as f:
lines = f.read().split('\n')[:-1]
pronunciation_variants_all = [line.split('\t') for line in lines]
word_list = np.delete(word_list, [0], 0) # remove 'Oog'
for word in word_list:
# load pronunciation variant of the word.
pronunciation_variants = []
for line in pronunciation_variants_all:
if line[0] == word.lower():
pronunciation_variants.append(line[1].replace(' ', ''))
# see only words which appears in top 3.
result_ = result[result['word'].str.match(word)]
result_tolerant = pd.DataFrame({
'filename': [''],
'word': [''],
'xsampa': [''],
'ipa': [''],
'prediction': [''],
'match': ['']})
for i in range(0, len(result_)):
line = result_.iloc[i]
# make a list of all possible pronunciation variants of ipa description.
# i.e. possible answers from forced alignment.
ipa = line['ipa']
pronvar_list = [ipa]
pronvar_list_ = am_func.fame_pronunciation_variant(ipa)
if not pronvar_list_ is None:
pronvar_list += list(pronvar_list_)
# only focus on pronunciations which can be estimated from ipa.
if len(set(pronvar_list) & set(pronunciation_variants)) > 0:
if line['prediction'] in pronvar_list:
ismatch = True
else:
ismatch = False
line_df = pd.DataFrame(result_.iloc[i]).T
df_idx = line_df.index[0]
result_tolerant_ = pd.Series([line_df.loc[df_idx, 'filename'],
line_df.loc[df_idx, 'word'],
line_df.loc[df_idx, 'xsampa'],
line_df.loc[df_idx, 'ipa'],
line_df.loc[df_idx, 'prediction'],
ismatch],
index=['filename', 'word', 'xsampa', 'ipa', 'prediction', 'match'],
name=df_idx)
result_tolerant = result_tolerant.append(result_tolerant_)
# remove the first entry (dummy)
result_tolerant = result_tolerant.drop(0, axis=0)
total_num = len(result_)
valid_num = len(result_tolerant)
match_num = np.sum(result_tolerant['match'])
print("word '{0}': {1}/{2} ({3:.2f} %) originally {4}".format(word, match_num, valid_num, match_num/valid_num*100, total_num))
f_result.write("{0},{1},{2},{3},{4}\n".format(word, total_num, valid_num, match_num, match_num/valid_num*100))
f_result.close()
## output confusion matrix
#cm = confusion_matrix(result_['ipa'], result_['prediction'])
#plt.figure()
#plot_confusion_matrix(cm, classes=pronunciation_variants, normalize=False)
#plt.savefig(result_dir + '\\cm_' + word + '.png')

14
acoustic_model/phoneset/fame_asr.py

@ -68,14 +68,21 @@ phoneset = [
# the phones which seldom occur are replaced with another more popular phones.
# replacements are based on the advice from Martijn Wieling.
reduction_key = {
'y':'i:', 'e':'e:', 'ə:':'ɛ:', 'r:':'r', 'ɡ':'g'
'y':'i:', 'e':'e:', 'ə:':'ɛ:', 'r:':'r', 'ɡ':'g',
# aki added because this is used in stimmen_project.
'ɔ̈:':'ɔ:'
}
# already removed beforehand in phoneset. Just to be sure.
phones_to_be_removed = ['ú', 's:', 'ɔ̈:']
phones_to_be_removed = ['ú', 's:']
def phone_reduction(phones):
"""
Args:
phones (list): list of phones.
"""
return [reduction_key.get(i, i) for i in phones
if not i in phones_to_be_removed]
phoneset_short = list(set(phone_reduction(phoneset)))
phoneset_short.sort()
@ -96,7 +103,8 @@ translation_key_asr2htk = {
'ŋ': 'ng',
# refer to Xsampa.
'ɔ': 'O', 'ɔ:': 'O:', 'ɔ̈': 'Oe',
'ɔ': 'O', 'ɔ:': 'O:', 'ɔ̈': 'Oe',
#'ɔ̈:': 'O:', # does not appear in FAME, but used in stimmen.
'ɛ': 'E', 'ɛ:': 'E:',
'ɪ': 'I', 'ɪ:': 'I:',

22
acoustic_model/stimmen_functions.py

@ -81,3 +81,25 @@ def add_row_asr(df):
for index, row in df.iterrows():
asr.append(fame_functions.ipa2asr(row['ipa']))
return df.assign(asr=asr)
def load_pronunciations(WORD, htk_dic):
""" load pronunciation variants from HTK dic file.
Args:
WORD (str): word in capital letters.
htk_dic (path): HTK dict file.
Returns:
(pronunciations) (list): pronunciation variants of WORD.
Notes:
Because this function loads all contents from htk_dic file,
it is not recommended to use for large lexicon.
"""
with open(htk_dic) as f:
lines = f.read().replace(' sil', '')
lines = lines.split('\n')
return [' '.join(line.split(' ')[1:])
for line in lines if line.split(' ')[0]==WORD]

18
acoustic_model/stimmen_test.py

@ -2,8 +2,9 @@ import os
os.chdir(r'C:\Users\Aki\source\repos\acoustic_model\acoustic_model')
import sys
import shutil
from collections import Counter
#import numpy as np
import numpy as np
import pandas as pd
import defaultfiles as default
@ -62,3 +63,18 @@ for ipa in df['ipa']:
if ':' in ipa_splitted:
print(ipa_splitted)
## check pronunciation variants
df_clean = stimmen_functions.load_transcriptions_clean(stimmen_test_dir)
df_clean = stimmen_functions.add_row_asr(df_clean)
df_clean = stimmen_functions.add_row_htk(df_clean)
for word in word_list:
#word = word_list[1]
df_ = df_clean[df_clean['word']==word]
c = Counter(df_['htk'])
pronunciations = dict()
for key, value in zip(c.keys(), c.values()):
if value > 3:
pronunciations[key] = value
print(pronunciations)
Loading…
Cancel
Save