triphone training is added.

acoustic_model/fame_functions.py

@@ -345,6 +345,7 @@ def fix_lexicon(lexicon_file):
 
 	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.
 	#lex.to_csv(lexicon_file, index=False, header=False, encoding="utf-8", sep=' ', quoting=csv.QUOTE_NONE, escapechar='\\')
 	lex.to_csv(lexicon_file, index=False, header=False, sep='\t', encoding='utf-8')

acoustic_model/fame_hmm.py

@@ -25,11 +25,11 @@ make_label		  = 0 # it takes roughly 4800 sec on Surface pro 2.
 make_mlf		  = 0
 extract_features  = 0
 flat_start		  = 0
-train_model_without_sp = 0
+train_monophone_without_sp = 0
 add_sp = 0
-train_model_with_re_aligned_mlf = 0
+train_monophone_with_re_aligned_mlf = 0
-train_triphone = 1
+train_triphone = 0
-
+train_triphone_tied = 1
 
 
 # pre-defined values.
@@ -46,16 +46,18 @@ lexicon_oov = os.path.join(default.fame_dir, 'lexicon', 'lex.oov')
 config_dir = os.path.join(default.htk_dir, 'config')
 
 model_dir     = os.path.join(default.htk_dir, 'model')
-model0_dir    = os.path.join(model_dir, 'hmm0')
+model_mono0_dir    = os.path.join(model_dir, 'mono0')
-model1_dir    = os.path.join(model_dir, 'hmm1')
+model_mono1_dir    = os.path.join(model_dir, 'mono1')
-model1sp_dir  = os.path.join(model_dir, 'hmm1sp')
+model_mono1sp_dir  = os.path.join(model_dir, 'mono1sp')
-model1sp2_dir = os.path.join(model_dir, 'hmm1sp2')
+model_mono1sp2_dir = os.path.join(model_dir, 'mono1sp2')
+model_tri1_dir  = os.path.join(model_dir, 'tri1')
 
 # 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')
+#lexicon_htk_with_sp = os.path.join(lexicon_dir, 'lex_with_sp.htk')
 
 feature_dir = os.path.join(default.htk_dir, 'mfc')
 fh.make_new_directory(feature_dir, existing_dir='leave')
@@ -68,6 +70,7 @@ fh.make_new_directory(label_dir, existing_dir='leave')
 ## 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_with_sp = os.path.join(label_dir, 'train_phone_with_sp.mlf')
 mlf_file_train_aligned = os.path.join(label_dir, 'train_phone_aligned.mlf')
 hcompv_scp_train_updated = hcompv_scp_train.replace('.scp', '_updated.scp')
 
@@ -97,8 +100,17 @@ if make_lexicon:
 	# 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))
 
+	## add sp to the end of each line.
+	#print('>>> adding sp...')
+	#with open(lexicon_htk) as f:
+	#	lines = f.read().split('\n')
+	#lines = [line + ' sp' for line in lines]
+	#with open(lexicon_htk_with_sp, 'wb') as f:
+	#	f.write(bytes('\n'.join(lines), 'ascii'))
+
+	print("elapsed time: {}".format(time.time() - timer_start))
+	
 
 ## intialize the instance for HTK.
 chtk = pyhtk.HTK(config_dir, fame_asr.phoneset_htk, lexicon_htk, feature_size)
@@ -164,12 +176,15 @@ if make_mlf:
 		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')
+		mlf_phone_with_sp = os.path.join(label_dir, dataset + '_phone_with_sp.mlf')
 
 		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)
+		chtk.mlf_word2phone(mlf_phone, mlf_word, with_sp=False)
+		chtk.mlf_word2phone(mlf_phone_with_sp, mlf_word, with_sp=True)
 		
+
 	print("elapsed time: {}".format(time.time() - timer_start))
 
 
@@ -224,33 +239,33 @@ if extract_features:
 if flat_start:
 	timer_start = time.time()
 	print('==== flat start ====')
-	fh.make_new_directory(model0_dir, existing_dir='leave')
+	fh.make_new_directory(model_mono0_dir, existing_dir='leave')
 
-	chtk.flat_start(hcompv_scp_train, model0_dir)
+	chtk.flat_start(hcompv_scp_train, model_mono0_dir)
 
 	# create macros.
-	vFloors = os.path.join(model0_dir, 'vFloors')
+	vFloors = os.path.join(model_mono0_dir, 'vFloors')
 	if os.path.exists(vFloors):
 		chtk.create_macros(vFloors)
 
 	# allocate mean & variance to all phones in the phone list
 	print('>>> allocating mean & variance to all phones in the phone list...')
 	chtk.create_hmmdefs(
-		os.path.join(model0_dir, proto_name),
+		os.path.join(model_mono0_dir, proto_name),
-	    os.path.join(model0_dir, 'hmmdefs')
+	    os.path.join(model_mono0_dir, 'hmmdefs')
 		)
 	
 	print("elapsed time: {}".format(time.time() - timer_start))
 
 
 ## ======================= train model without short pause =======================
-if train_model_without_sp:
+if train_monophone_without_sp:
-	print('==== train model without sp ====')
+	print('==== train monophone without sp ====')
 	
 	timer_start = time.time()
 	niter = chtk.re_estimation_until_saturated(
-		model1_dir,
+		model_mono1_dir,
-		model0_dir, improvement_threshold, hcompv_scp_train, 
+		model_mono0_dir, improvement_threshold, hcompv_scp_train, 
 		os.path.join(htk_stimmen_dir, 'mfc'), 
 		'mfc', 
 		os.path.join(htk_stimmen_dir, 'word_lattice.ltc'), 
@@ -270,32 +285,34 @@ if add_sp:
 
 	# make model with sp.
 	print('>>> adding sp state to the last model in the previous step...')
-	fh.make_new_directory(model1sp_dir, existing_dir='leave')
+	fh.make_new_directory(model_mono1sp_dir, existing_dir='leave')
-	niter = chtk.get_niter_max(model1_dir)
+	niter = chtk.get_niter_max(model_mono1_dir)
-	modeln_dir_pre = os.path.join(model1_dir, 'iter'+str(niter))
+	modeln_dir_pre = os.path.join(model_mono1_dir, 'iter'+str(niter))
-	modeln_dir	   = os.path.join(model1sp_dir, 'iter0')
+	modeln_dir	   = os.path.join(model_mono1sp_dir, 'iter0')
+
+	#hmmdefs_pre = os.path.join(modeln_dir_pre, 'hmmdefs')
 	chtk.add_sp(modeln_dir_pre, modeln_dir)
 	print("elapsed time: {}".format(time.time() - timer_start))
 
 	niter = chtk.re_estimation_until_saturated(
-		model1sp_dir, modeln_dir, improvement_threshold, hcompv_scp_train, 
+		model_mono1sp_dir, modeln_dir, improvement_threshold, hcompv_scp_train, 
 		os.path.join(htk_stimmen_dir, 'mfc'), 
 		'mfc', 
 		os.path.join(htk_stimmen_dir, 'word_lattice.ltc'), 
-		mlf_file=mlf_file_train, 
+		mlf_file=mlf_file_train_with_sp, 
 		lexicon_file=os.path.join(htk_stimmen_dir, 'lexicon_recognition.dic'), 
 		model_type='monophone_with_sp'
 		)
 		
 
 ## ======================= train model with re-aligned mlf =======================
-if train_model_with_re_aligned_mlf:
+if train_monophone_with_re_aligned_mlf:
-	print('==== traina model with re-aligned mlf ====')
+	print('==== traina monophone with re-aligned mlf ====')
 
 	print('>>> re-aligning the training data... ')
 	timer_start = time.time()
-	niter = chtk.get_niter_max(model1sp_dir)
+	niter = chtk.get_niter_max(model_mono1sp_dir)
-	modeln_dir = os.path.join(model1sp_dir, 'iter'+str(niter))
+	modeln_dir = os.path.join(model_mono1sp_dir, 'iter'+str(niter))
 	chtk.make_aligned_label(
 		os.path.join(modeln_dir, 'macros'),
 		os.path.join(modeln_dir, 'hmmdefs'), 
@@ -306,18 +323,18 @@ if train_model_with_re_aligned_mlf:
 	print('>>> updating the script file... ')
 	chtk.update_script_file(
 		mlf_file_train_aligned, 
-		mlf_file_train, 
+		mlf_file_train_with_sp, 
 		hcompv_scp_train, 
 		hcompv_scp_train_updated)
 	print("elapsed time: {}".format(time.time() - timer_start))
 
 	print('>>> re-estimation... ')
 	timer_start = time.time()
-	fh.make_new_directory(model1sp2_dir, existing_dir='leave')
+	fh.make_new_directory(model_mono1sp2_dir, existing_dir='leave')
-	niter = chtk.get_niter_max(model1sp_dir)
+	niter = chtk.get_niter_max(model_mono1sp_dir)
 	niter = chtk.re_estimation_until_saturated(
-		model1sp2_dir, 
+		model_mono1sp2_dir, 
-		os.path.join(model1sp_dir, 'iter'+str(niter)), 
+		os.path.join(model_mono1sp_dir, 'iter'+str(niter)), 
 		improvement_threshold, 
 		hcompv_scp_train_updated, 
 		os.path.join(htk_stimmen_dir, 'mfc'), 
@@ -332,25 +349,68 @@ if train_model_with_re_aligned_mlf:
 
 ## ======================= train triphone =======================
 if train_triphone:
-	model_out_dir = os.path.join(model_dir, 'hmm1_tri', 'iter1')
+	print('==== traina triphone model ====')
+	#model_out_dir = os.path.join(model_dir, 'hmm1_tri', 'iter1')
 
-	triphonelist_txt = os.path.join(config_dir, 'triphonelist_txt')
+	triphonelist_txt = os.path.join(config_dir, 'triphonelist.txt')
 	triphone_mlf = os.path.join(default.htk_dir, 'label', 'train_triphone.mlf')
 
+	print('>>> making triphone list... ')
 	chtk.make_triphonelist(
 		triphonelist_txt, 
 		triphone_mlf, 
 		mlf_file_train_aligned)
 
-	#run_command([
+	print('>>> making triphone header... ')
-	#	'HERest', '-B', 
+	chtk.make_tri_hed(
-	#	'-C', config_train,
+		os.path.join(config_dir, 'mktri.hed')
-	#	'-I', triphone_mlf,
+		)
-	#	'-t', '250.0', '150.0', '1000.0',
+
-	#	'-s', 'stats' 
+	print('>>> init triphone model... ')
-	#	'-S', hcompv_scp_train,
+	niter = chtk.get_niter_max(model_mono1sp2_dir)
-	#	'-H', macros,
+	fh.make_new_directory(os.path.join(model_tri1_dir, 'iter0'), existing_dir='leave')
-	#	'-H', hmmdefs,
+	chtk.init_triphone(
-	#	'-M', model_out_dir, 
+		os.path.join(model_mono1sp2_dir, 'iter'+str(niter)),
-	#	os.path.join(config_dir, 'triphonelist.txt')
+		os.path.join(model_tri1_dir, 'iter0')
-	#])
+		)
+
+	print('>>> re-estimation... ')
+	# I wanted to train until satulated:
+	# 	#niter = chtk.re_estimation_until_saturated(
+	#	model_tri1_dir, 
+	#	os.path.join(model_tri1_dir, 'iter0'), 
+	#	improvement_threshold, 
+	#	hcompv_scp_train_updated, 
+	#	os.path.join(htk_stimmen_dir, 'mfc'), 
+	#	'mfc', 
+	#	os.path.join(htk_stimmen_dir, 'word_lattice.ltc'), 
+	#	mlf_file=triphone_mlf, 
+	#	lexicon_file=os.path.join(htk_stimmen_dir, 'lexicon_recognition.dic'), 
+	#	model_type='triphone'
+	#	)
+	#
+	# but because the data size is limited, some triphone cannot be trained and received the error:
+	#   ERROR [+8231]  GetHCIModel: Cannot find hmm [i:-]r[+???]
+	# therefore only two times re-estimation is performed.
+	output_dir = model_tri1_dir
+
+	for niter in range(1, 4):
+		hmm_n = 'iter' + str(niter)
+		hmm_n_pre = 'iter' + str(niter-1)
+		_modeln_dir	    = os.path.join(output_dir, hmm_n)
+		_modeln_dir_pre = os.path.join(output_dir, hmm_n_pre) 
+		
+		fh.make_new_directory(_modeln_dir, 'leave')
+		chtk.re_estimation(
+			os.path.join(_modeln_dir_pre, 'hmmdefs'), 
+			_modeln_dir,
+			hcompv_scp_train_updated,
+			mlf_file=triphone_mlf,
+			macros=os.path.join(_modeln_dir_pre, 'macros'),
+			model_type='triphone')
+
+
+## ======================= train triphone =======================
+if train_triphone_tied:
+	print('==== traina tied-state triphone ====')
+