sp is added to the model.

acoustic_model/fame_hmm.py

@@ -25,13 +25,17 @@ make_label		  = 0 # it takes roughly 4800 sec on Surface pro 2.
 make_htk_files    = 0
 extract_features  = 0
 flat_start		  = 0
-train_model_without_sp = 1
+train_model_without_sp = 0
+add_sp = 0
+train_model_with_sp    = 0
+
 
 
 # pre-defined values.
 
 dataset_list = ['devel', 'test', 'train']
 hmmdefs_name = 'hmmdefs'
+proto_name   = 'proto39'
 
 lexicon_asr = os.path.join(default.fame_dir, 'lexicon', 'lex.asr')
 lexicon_oov = os.path.join(default.fame_dir, 'lexicon', 'lex.oov')
@@ -41,7 +45,8 @@ 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')
-prototype    = os.path.join(config_dir, 'proto39')
+sil_hed		 = os.path.join(config_dir, 'sil.hed')
+prototype    = os.path.join(config_dir, proto_name)
 
 model_dir    = os.path.join(default.htk_dir, 'model')
 
@@ -55,6 +60,7 @@ 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')
 
 feature_dir = os.path.join(default.htk_dir, 'mfc')
 if not os.path.exists(feature_dir):
@@ -66,6 +72,13 @@ label_dir = os.path.join(default.htk_dir, 'label')
 if not os.path.exists(label_dir):
 	os.makedirs(label_dir)
 
+## training
+hcompv_scp_train = os.path.join(tmp_dir, 'train.scp')
+mlf_file_train   = os.path.join(label_dir, 'train_phone.mlf')
+
+## train without sp
+niter_max = 10
+
 
 ## ======================= make lexicon for HTK =======================
 if make_lexicon:
@@ -95,6 +108,7 @@ if make_lexicon:
 
 ## ======================= 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))
@@ -216,45 +230,111 @@ if extract_features:
 
 ## ======================= flat start monophones =======================
 if flat_start:
-	hcompv_scp = os.path.join(tmp_dir, 'test.scp')
-
 	timer_start = time.time()
 	print('==== flat start ====')
-	pyhtk.flat_start(config_train, hcompv_scp, model0_dir, prototype)
+	pyhtk.flat_start(config_train, hcompv_scp_train, model0_dir, prototype)
 
 	# allocate mean & variance to all phones in the phone list
+	print('>>> allocating mean & variance to all phones in the phone list...')
 	pyhtk.create_hmmdefs(
-		os.path.join(model0_dir, 'proto39'),
+		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 <MFCC_0_D_A> <VecSize> 39\n' + lines, 'ascii'))
+	
 	print("elapsed time: {}".format(time.time() - timer_start))
 
 
-## ======================= estimate monophones =======================
+## ======================= train model without short pause =======================
 if train_model_without_sp:
-	hcompv_scp = os.path.join(tmp_dir, 'test.scp')
+	fh.make_new_directory(model1_dir)
-	mlf_file = os.path.join(label_dir, 'test_phone.mlf')
-	output_dir = os.path.join(model_dir, 'hmm1')
-	fh.make_new_directory(output_dir)
 
 	print('==== train model without sp ====')
-	if not os.path.exists(os.path.join(output_dir, 'iter0')):
+	if not os.path.exists(os.path.join(model1_dir, 'iter0')):
-		shutil.copytree(model0_dir, os.path.join(output_dir, 'iter0'))
+		shutil.copytree(model0_dir, os.path.join(model1_dir, 'iter0'))
-	niter = 1
+	for niter in range(1, niter_max):
-	for niter in range(1, 5):
 		timer_start = time.time()
 		hmm_n = 'iter' + str(niter)
 		hmm_n_pre = 'iter' + str(niter-1)
-		modeln_dir	   = os.path.join(output_dir, hmm_n)
+		modeln_dir	   = os.path.join(model1_dir, hmm_n)
-		modeln_dir_pre = os.path.join(output_dir, hmm_n_pre) 
+		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, 'proto39'),
+			os.path.join(modeln_dir_pre, 'macros'),
 			os.path.join(modeln_dir_pre, hmmdefs_name), 
 			modeln_dir,
-			hcompv_scp, phonelist_txt,
+			hcompv_scp_train, phonelist_txt,
-			mlf_file=mlf_file)
+			mlf_file=mlf_file_train)
+		print("elapsed time: {}".format(time.time() - timer_start))
+
+
+## ======================= adding sp to the model =======================
+if add_sp:
+	print('==== adding sp to the model ====')
+
+	# make model with sp.
+	print('>>> modifying the last model in the previous step...')
+	modeln_dir_pre = os.path.join(model1_dir, 'iter'+str(niter_max-1))
+	modeln_dir = modeln_dir_pre.replace('iter' + str(niter_max-1), 'iter' + str(niter_max))
+	fh.make_new_directory(modeln_dir)
+	shutil.copy(
+		os.path.join(modeln_dir_pre, 'macros'), 
+		os.path.join(modeln_dir, 'macros'))
+	shutil.copy(
+		os.path.join(modeln_dir_pre, hmmdefs_name), 
+		os.path.join(modeln_dir, hmmdefs_name))
+
+	## ======================= 
+	## manually make changes to modeln_dir/hmmdefs 
+	## ======================= 
+	# add states 'sil'.
+	# http://www.f.waseda.jp/yusukekondo/htk.html#flat_start_estimation
+	#shutil.copy(
+	#	os.path.join(model_dir, 'hmmdefs.txt'), 
+	#	os.path.join(modeln_dir, hmmdefs_name))
+
+	#hmmdefs_file_pre = os.path.join(modeln_dir_pre, hmmdefs_name)
+	hmmdefs_file = os.path.join(modeln_dir, hmmdefs_name)
+	macros_file  = os.path.join(modeln_dir, 'macros')
+	#with open(hmmdefs_file_pre) as f:
+	#	lines = f.read()
+	#lines_ = lines.split('~h ')
+	#sil_model = [line for line in lines_ if line.split('\n')[0].replace('"', '') == 'sil'][0]
+	
+	# update hmmdefs and macros.
+	print('>>> updating hmmdefs and macros...')
+	modeln_dir_pre = modeln_dir
+	modeln_dir = modeln_dir.replace('iter' + str(niter_max), 'iter' + str(niter_max+1))
+	fh.make_new_directory(modeln_dir)
+	pyhtk.include_sil_in_hmmdefs(macros_file, hmmdefs_file, modeln_dir, sil_hed, phonelist_txt)
+
+
+## ======================= 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):
+		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, 'macros'),
+			os.path.join(modeln_dir_pre, hmmdefs_name), 
+			modeln_dir,
+			hcompv_scp_train, phonelist_txt,
+			mlf_file=mlf_file_train)
 		print("elapsed time: {}".format(time.time() - timer_start))