diff --git a/.vs/acoustic_model/v15/.suo b/.vs/acoustic_model/v15/.suo
index 1238af1..712c255 100644
Binary files a/.vs/acoustic_model/v15/.suo and b/.vs/acoustic_model/v15/.suo differ
diff --git a/acoustic_model/acoustic_model.pyproj b/acoustic_model/acoustic_model.pyproj
index ebdbadc..715d6c0 100644
--- a/acoustic_model/acoustic_model.pyproj
+++ b/acoustic_model/acoustic_model.pyproj
@@ -4,7 +4,7 @@
2.0
4d8c8573-32f0-4a62-9e62-3ce5cc680390
.
- htk_vs_kaldi.py
+ fame_hmm.py
.
diff --git a/acoustic_model/fame_functions.py b/acoustic_model/fame_functions.py
index 9d3992b..9ca7e0d 100644
--- a/acoustic_model/fame_functions.py
+++ b/acoustic_model/fame_functions.py
@@ -12,6 +12,10 @@ import defaultfiles as default
import convert_phoneset
from phoneset import fame_ipa, fame_asr
+sys.path.append(default.toolbox_dir)
+from htk import pyhtk
+
+
#def read_fileFA(fileFA):
# """
# read the result file of HTK forced alignment.
@@ -371,4 +375,25 @@ def ipa2htk(ipa):
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)
\ No newline at end of file
+ return ''.join(htk_splitted)
+
+
+def performance_on_stimmen(stimmen_dir, hmmdefs):
+ #hmmdefs = r'c:\OneDrive\Research\rug\experiments\acoustic_model\fame\htk\model_\hmm1\iter20\hmmdefs'
+ #stimmen_dir = r'c:\OneDrive\Research\rug\experiments\acoustic_model\fame\htk\stimmen'
+ lattice_file = os.path.join(stimmen_dir, 'word_lattice.ltc')
+ hvite_scp = os.path.join(stimmen_dir, 'hvite.scp')
+ #fh.make_filelist(os.path.join(stimmen_dir, 'mfc'), hvite_scp, file_type='mfc')
+ hresult_scp = os.path.join(stimmen_dir, 'hresult.scp')
+ #fh.make_filelist(os.path.join(stimmen_dir, 'mfc'), hresult_scp, file_type='rec')
+ lexicon_file = os.path.join(stimmen_dir, 'lexicon_recognition.dic')
+ chtk = pyhtk.HTK(config_dir, fame_asr.phoneset_htk, lexicon_file)
+
+ result = chtk.recognition(
+ lattice_file,
+ hmmdefs,
+ hvite_scp
+ )
+ per_sentence, per_word = chtk.calc_recognition_performance(hresult_scp)
+
+ return per_sentence['accuracy']
\ No newline at end of file
diff --git a/acoustic_model/fame_hmm.py b/acoustic_model/fame_hmm.py
index 7228c00..4e2f430 100644
--- a/acoustic_model/fame_hmm.py
+++ b/acoustic_model/fame_hmm.py
@@ -22,30 +22,27 @@ from htk import pyhtk
# procedure
make_lexicon = 0
make_label = 0 # it takes roughly 4800 sec on Surface pro 2.
-make_htk_files = 0
+make_mlf = 0
extract_features = 0
flat_start = 0
train_model_without_sp = 0
add_sp = 0
train_model_with_sp = 0
-train_model_with_sp_align_mlf = 1
+train_model_with_sp_align_mlf = 0
+train_triphone = 0
# pre-defined values.
-
dataset_list = ['devel', 'test', 'train']
hmmdefs_name = 'hmmdefs'
-proto_name = 'proto39'
+proto_name = 'proto'
lexicon_asr = os.path.join(default.fame_dir, 'lexicon', 'lex.asr')
lexicon_oov = os.path.join(default.fame_dir, 'lexicon', 'lex.oov')
config_dir = os.path.join(default.htk_dir, 'config')
-config_hcopy = os.path.join(config_dir, 'config.HCopy')
-config_train = os.path.join(config_dir, 'config.train')
-global_ded = os.path.join(config_dir, 'global.ded')
-mkphones_led = os.path.join(config_dir, 'mkphones.led')
+
sil_hed = os.path.join(config_dir, 'sil.hed')
prototype = os.path.join(config_dir, proto_name)
@@ -53,25 +50,20 @@ model_dir = os.path.join(default.htk_dir, 'model')
# directories / files to be made.
-
lexicon_dir = os.path.join(default.htk_dir, 'lexicon')
lexicon_htk_asr = os.path.join(lexicon_dir, 'lex.htk_asr')
lexicon_htk_oov = os.path.join(lexicon_dir, 'lex.htk_oov')
lexicon_htk = os.path.join(lexicon_dir, 'lex.htk')
-phonelist_txt = os.path.join(config_dir, 'phonelist.txt')
-model0_dir = os.path.join(model_dir, 'hmm0')
-model1_dir = os.path.join(model_dir, 'hmm1')
+
+#model1_dir = os.path.join(model_dir, 'hmm1')
feature_dir = os.path.join(default.htk_dir, 'mfc')
-if not os.path.exists(feature_dir):
- os.makedirs(feature_dir)
+fh.make_new_directory(feature_dir, existing_dir='leave')
tmp_dir = os.path.join(default.htk_dir, 'tmp')
-if not os.path.exists(tmp_dir):
- os.makedirs(tmp_dir)
+fh.make_new_directory(tmp_dir, existing_dir='leave')
label_dir = os.path.join(default.htk_dir, 'label')
-if not os.path.exists(label_dir):
- os.makedirs(label_dir)
+fh.make_new_directory(label_dir, existing_dir='leave')
## training
hcompv_scp_train = os.path.join(tmp_dir, 'train.scp')
@@ -98,20 +90,21 @@ if make_lexicon:
# therefore there is no overlap between lex_asr and lex_oov.
fame_functions.combine_lexicon(lexicon_htk_asr, lexicon_htk_oov, lexicon_htk)
- ## =======================
- ## manually make changes to the pronunciation dictionary and save it as lex.htk
- ## =======================
+ ## fixing the lexicon for HTK.
# (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
+ # http://electroblaze.blogspot.nl/2013/03/understanding-htk-error-messages.html
print('>>> fixing the lexicon...')
fame_functions.fix_lexicon(lexicon_htk)
print("elapsed time: {}".format(time.time() - timer_start))
+## intialize the instance for HTK.
+chtk = pyhtk.HTK(config_dir, fame_asr.phoneset_htk, lexicon_htk)
+
+
## ======================= make label files =======================
if make_label:
- # train_2002_gongfansaken_10347.lab is empty. should be removed.
for dataset in dataset_list:
timer_start = time.time()
print("==== making label files on dataset {}".format(dataset))
@@ -120,7 +113,7 @@ if make_label:
wav_dir_ = os.path.join(default.fame_dir, 'fame', 'wav', dataset)
label_dir_ = os.path.join(label_dir, dataset)
dictionary_file = os.path.join(label_dir_, 'temp.dic')
- fh.make_new_directory(label_dir_)
+ fh.make_new_directory(label_dir_, existing_dir='leave')
# list of scripts
with open(script_list, "rt", encoding="utf-8") as fin:
@@ -135,56 +128,48 @@ if make_label:
sentence_htk = fame_functions.word2htk(sentence)
wav_file = os.path.join(wav_dir_, filename + '.wav')
- if os.path.exists(wav_file) and pyhtk.can_be_ascii(sentence_htk) == 0:
- if pyhtk.create_dictionary_without_log(
- sentence_htk, global_ded, dictionary_file, lexicon_htk) == 0:
+ if os.path.exists(wav_file) and chtk.can_be_ascii(sentence_htk) == 0:
+ if chtk.get_number_of_missing_words(
+ sentence_htk, dictionary_file) == 0:
# when the file name is too long, HDMan command does not work.
# therefore first temporary dictionary_file is made, then renamed.
shutil.move(dictionary_file, os.path.join(label_dir_, filename + '.dic'))
label_file = os.path.join(label_dir_, filename + '.lab')
- pyhtk.create_label_file(sentence_htk, label_file)
+ chtk.create_label_file(sentence_htk, label_file)
else:
os.remove(dictionary_file)
+
print("elapsed time: {}".format(time.time() - timer_start))
-## ======================= make other required files =======================
-if make_htk_files:
+## ======================= make master label files =======================
+if make_mlf:
timer_start = time.time()
- print("==== making files required for HTK ====")
+ print("==== making master label files ====")
- print(">>> making a phonelist...")
- pyhtk.create_phonelist_file(fame_asr.phoneset_htk, phonelist_txt)
+ # train_2002_gongfansaken_10347.lab is empty. should be removed.
+ empty_lab_file = os.path.join(label_dir, 'train', 'train_2002_gongfansaken_10347.lab')
+ empty_dic_file = empty_lab_file.replace('.lab', '.dic')
+
+ if os.path.exists(empty_lab_file):
+ os.remove(empty_lab_file)
+ if os.path.exists(empty_dic_file):
+ os.remove(empty_dic_file)
for dataset in dataset_list:
- wav_dir_ = os.path.join(default.fame_dir, 'fame', 'wav', dataset)
+ #wav_dir_ = os.path.join(default.fame_dir, 'fame', 'wav', dataset)
feature_dir_ = os.path.join(feature_dir, dataset)
label_dir_ = os.path.join(label_dir, dataset)
mlf_word = os.path.join(label_dir, dataset + '_word.mlf')
mlf_phone = os.path.join(label_dir, dataset + '_phone.mlf')
- #print(">>> making a script file for {}...".format(dataset))
- #listdir = glob.glob(os.path.join(wav_dir_, '*.dic'))
- #mfc_list = [filename.replace(wav_dir_, feature_dir_).replace('.dic', '.mfc') for filename in listdir]
- #hcompv_scp = os.path.join(tmp_dir, dataset + '.scp')
- #with open(hcompv_scp, 'wb') as f:
- # f.write(bytes('\n'.join(mfc_list) + '\n', 'ascii'))
-
- print(">>> making a mlf file for {}...".format(dataset))
- lab_list = glob.glob(os.path.join(label_dir_, '*.lab'))
- with open(mlf_word, 'wb') as fmlf:
- fmlf.write(bytes('#!MLF!#\n', 'ascii'))
- for label_file in lab_list:
- filename = os.path.basename(label_file)
- fmlf.write(bytes('\"*/{}\"\n'.format(filename), 'ascii'))
- with open(label_file) as flab:
- lines = flab.read()
- fmlf.write(bytes(lines + '.\n', 'ascii'))
-
- print(">>> generating phone level transcription for {}...".format(dataset))
- pyhtk.mlf_word2phone(lexicon_htk, mlf_phone, mlf_word, mkphones_led)
- print("elapsed time: {}".format(time.time() - timer_start))
+ print(">>> generating a word level mlf file for {}...".format(dataset))
+ chtk.label2mlf(label_dir_, mlf_word)
+ print(">>> generating a phone level mlf file for {}...".format(dataset))
+ chtk.mlf_word2phone(mlf_phone, mlf_word)
+
+ print("elapsed time: {}".format(time.time() - timer_start))
## ======================= extract features =======================
@@ -196,7 +181,7 @@ if extract_features:
wav_dir_ = os.path.join(default.fame_dir, 'fame', 'wav', dataset)
label_dir_ = os.path.join(label_dir, dataset)
feature_dir_ = os.path.join(feature_dir, dataset)
- fh.make_new_directory(feature_dir_)
+ fh.make_new_directory(feature_dir_, existing_dir='delete')
# a script file for HCopy
print(">>> making a script file for HCopy...")
@@ -212,12 +197,15 @@ if extract_features:
os.path.join(wav_dir_, os.path.basename(lab_file).replace('.lab', '.wav')) + '\t'
+ os.path.join(feature_dir_, os.path.basename(lab_file).replace('.lab', '.mfc'))
for lab_file in lab_list]
+
+ if os.path.exists(empty_mfc_file):
+ os.remove(empty_mfc_file)
with open(hcopy_scp.name, 'wb') as f:
f.write(bytes('\n'.join(feature_list), 'ascii'))
# extract features.
print(">>> extracting features on {}...".format(dataset))
- pyhtk.wav2mfc(config_hcopy, hcopy_scp.name)
+ chtk.wav2mfc(hcopy_scp.name)
os.remove(hcopy_scp.name)
# make hcompv.scp.
@@ -235,21 +223,18 @@ if extract_features:
if flat_start:
timer_start = time.time()
print('==== flat start ====')
- pyhtk.flat_start(config_train, hcompv_scp_train, model0_dir, prototype)
+ feature_size = 39
+ model0_dir = os.path.join(model_dir, 'hmm0')
+ fh.make_new_directory(model0_dir, existing_dir='leave')
+
+ chtk.flat_start(hcompv_scp_train, model0_dir, feature_size)
# allocate mean & variance to all phones in the phone list
print('>>> allocating mean & variance to all phones in the phone list...')
- pyhtk.create_hmmdefs(
+ chtk.create_hmmdefs(
os.path.join(model0_dir, proto_name),
- os.path.join(model0_dir, 'hmmdefs'),
- phonelist_txt)
-
- # make macros
- print('>>> making macros...')
- with open(os.path.join(model0_dir, 'vFloors')) as f:
- lines = f.read()
- with open(os.path.join(model0_dir, 'macros'), 'wb') as f:
- f.write(bytes('~o 39\n' + lines, 'ascii'))
+ os.path.join(model0_dir, 'hmmdefs')
+ )
print("elapsed time: {}".format(time.time() - timer_start))
@@ -362,4 +347,24 @@ if train_model_with_sp_align_mlf:
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))
\ No newline at end of file
+ print("elapsed time: {}".format(time.time() - timer_start))
+
+
+# train triphone.
+if train_triphone:
+ triphone_mlf = os.path.join(default.htk_dir, 'label', 'train_triphone.mlf')
+ macros = os.path.join(model_dir, 'hmm1_tri', 'iter0', 'macros')
+ hmmdefs = os.path.join(model_dir, 'hmm1_tri', 'iter0', 'hmmdefs')
+ model_out_dir = os.path.join(model_dir, 'hmm1_tri', 'iter1')
+ run_command([
+ 'HERest', '-B',
+ '-C', config_train,
+ '-I', triphone_mlf,
+ '-t', '250.0', '150.0', '1000.0',
+ '-s', 'stats'
+ '-S', hcompv_scp_train,
+ '-H', macros,
+ '-H', hmmdefs,
+ '-M', model_out_dir,
+ os.path.join(config_dir, 'triphonelist.txt')
+ ])
\ No newline at end of file
diff --git a/acoustic_model/htk_vs_kaldi.py b/acoustic_model/htk_vs_kaldi.py
index c35a42f..5297b79 100644
--- a/acoustic_model/htk_vs_kaldi.py
+++ b/acoustic_model/htk_vs_kaldi.py
@@ -53,7 +53,7 @@ from htk import pyhtk
# procedure
make_dic_file = 0
-make_HTK_files = 1
+make_HTK_files = 0
extract_features = 0
#make_htk_dict_files = 0
#do_forced_alignment_htk = 0
@@ -171,7 +171,7 @@ if make_HTK_files:
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'
+ label_string = 'SILENCE\n' + row['word'].upper() + '\nSILENCE\n'
f.write(bytes(label_string, 'ascii'))
@@ -249,7 +249,7 @@ with open(hresult_scp, 'wb') as f:
# calculate result
performance = np.zeros((1, 2))
-for niter in range(1, 50):
+for niter in range(50, 60):
output = pyhtk.recognition(
os.path.join(config_dir, 'config.rec'),
lattice_file,
@@ -265,6 +265,16 @@ for niter in range(1, 50):
+ #output = run_command_with_output([
+ # 'HVite', '-T', '1',
+ # '-C', config_rec,
+ # '-w', lattice_file,
+ # '-H', hmm,
+ # dictionary_file, phonelist_txt,
+ # '-S', HVite_scp
+ #])
+
+
## ======================= forced alignment using HTK =======================
if do_forced_alignment_htk:
diff --git a/acoustic_model/phoneset/fame_asr.py b/acoustic_model/phoneset/fame_asr.py
index 6165d5c..398d2b3 100644
--- a/acoustic_model/phoneset/fame_asr.py
+++ b/acoustic_model/phoneset/fame_asr.py
@@ -128,7 +128,11 @@ translation_key_word2htk = {
'ä': 'ao', 'ë': 'ee', 'ï': 'ie', 'ö': 'oe', 'ü': 'ue',
}
#[translation_key_word2htk.get(i, i) for i in not_in_ascii]
-
+#Stop: p, b, t, d, k, g
+#Nasal: m, n, ng(ŋ)
+#Fricative: s, z, f, v, h, x
+#Liquid: l, r
+#Vowel: a, a:, e:, i, i:, i_(i̯), o, o:, u, u:, u_(ṷ), oe(ö), oe:(ö:), ue(ü), ue:(ü:), O(ɔ), O:(ɔ:), Oe(ɔ̈), A(ə), E(ɛ), E:(ɛ:), I(ɪ), I:(ɪ:)
## the list of multi character phones.
diff --git a/acoustic_model/stimmen_test.py b/acoustic_model/stimmen_test.py
index 93546ca..f7911a7 100644
--- a/acoustic_model/stimmen_test.py
+++ b/acoustic_model/stimmen_test.py
@@ -77,4 +77,17 @@ for word in word_list:
for key, value in zip(c.keys(), c.values()):
if value > 3:
pronunciations[key] = value
- print(pronunciations)
\ No newline at end of file
+ print(pronunciations)
+
+
+monophone_mlf = os.path.join(default.htk_dir, 'label', 'train_phone_aligned.mlf')
+triphone_mlf = os.path.join(default.htk_dir, 'label', 'train_triphone.mlf')
+def filenames_in_mlf(file_mlf):
+ with open(file_mlf) as f:
+ lines_ = f.read().split('\n')
+ lines = [line for line in lines_ if len(line.split(' ')) == 1 and line != '.']
+ filenames = [line.replace('"', '').replace('*/', '') for line in lines[1:-1]]
+ return filenames
+filenames_mono = filenames_in_mlf(monophone_mlf)
+filenames_tri = filenames_in_mlf(triphone_mlf)
+