Code is cleaned up.

acoustic_model.sln

@@ -11,6 +11,7 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Solution Items", "Solution
 		..\forced_alignment\forced_alignment\convert_phone_set.py = ..\forced_alignment\forced_alignment\convert_phone_set.py
 		..\toolbox\evaluation.py = ..\toolbox\evaluation.py
 		..\forced_alignment\forced_alignment\forced_alignment.pyproj = ..\forced_alignment\forced_alignment\forced_alignment.pyproj
+		..\forced_alignment\forced_alignment\htk_dict.py = ..\forced_alignment\forced_alignment\htk_dict.py
 		..\forced_alignment\forced_alignment\lexicon.py = ..\forced_alignment\forced_alignment\lexicon.py
 		..\forced_alignment\forced_alignment\mlf.py = ..\forced_alignment\forced_alignment\mlf.py
 		..\forced_alignment\forced_alignment\pronunciations.py = ..\forced_alignment\forced_alignment\pronunciations.py

acoustic_model/acoustic_model_functions.py

@@ -38,7 +38,7 @@ def make_filelist(input_dir, output_txt):
 			fout.write(input_dir + '\\' + filename + '\n')
 
 
-def make_dic(word, pronvar_, fileDic, output_type):
+def make_htk_dict(word, pronvar_, fileDic, output_type):
     """
     make dict files which can be used for HTK.
     param word: target word.
@@ -98,8 +98,8 @@ def find_phone(lexicon_file, phone):
 	for line in lines:
 		line = line.split('\t')
 		if len(line) > 1:
-			pron = line[1]
+			pronunciation = line[1]
-			if phone in pron:
+			if phone in pronunciation:
 				extracted.append(line)
 	return extracted
 
@@ -149,3 +149,54 @@ def read_fileFA(fileFA):
             phones.append(line_split[2])
 
     return ' '.join(phones)
+
+
+def fame_pronunciation_variant(ipa):
+    ipa = ipa.replace('æ', 'ɛ')
+    ipa = ipa.replace('ɐ', 'a')
+    ipa = ipa.replace('ɑ', 'a')
+    ipa = ipa.replace('ɾ', 'r')
+    ipa = ipa.replace('ɹ', 'r') # ???
+    ipa = ipa.replace('ʁ', 'r')
+    ipa = ipa.replace('ʀ', 'r') # ???
+    ipa = ipa.replace('ʊ', 'u')
+    ipa = ipa.replace('χ', 'x')
+
+    pronvar_list = [ipa]
+    while 'ø:' in ' '.join(pronvar_list) or 'œ' in ' '.join(pronvar_list) or 'ɒ' in ' '.join(pronvar_list):
+        pronvar_list_ = []
+        for p in pronvar_list:
+            if 'ø:' in p:
+                pronvar_list_.append(p.replace('ø:', 'ö'))
+                pronvar_list_.append(p.replace('ø:', 'ö:'))
+            if 'œ' in p:
+                pronvar_list_.append(p.replace('œ', 'ɔ̈'))
+                pronvar_list_.append(p.replace('œ', 'ɔ̈:'))
+            if 'ɒ' in p:
+                pronvar_list_.append(p.replace('ɒ', 'ɔ̈'))
+                pronvar_list_.append(p.replace('ɒ', 'ɔ̈:'))
+        pronvar_list = np.unique(pronvar_list_)
+    return pronvar_list
+
+
+def make_fame2ipa_variants(fame):
+    fame = 'rɛös'
+    ipa = [fame]
+    ipa.append(fame.replace('ɛ', 'æ'))
+    ipa.append(fame.replace('a', 'ɐ'))
+    ipa.append(fame.replace('a', 'ɑ'))
+    ipa.append(fame.replace('r', 'ɾ'))
+    ipa.append(fame.replace('r', 'ɹ'))
+    ipa.append(fame.replace('r', 'ʁ'))
+    ipa.append(fame.replace('r', 'ʀ'))
+    ipa.append(fame.replace('u', 'ʊ'))
+    ipa.append(fame.replace('x', 'χ'))
+        
+    ipa.append(fame.replace('ö', 'ø:'))
+    ipa.append(fame.replace('ö:', 'ø:'))
+    ipa.append(fame.replace('ɔ̈', 'œ'))
+    ipa.append(fame.replace('ɔ̈:', 'œ'))
+    ipa.append(fame.replace('ɔ̈', 'ɒ'))
+    ipa.append(fame.replace('ɔ̈:', 'ɒ'))
+
+    return ipa

acoustic_model/defaultfiles.py

@@ -2,7 +2,8 @@ import os
 
 #default_hvite_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), 'data', 'htk', 'config.HVite')
 
-cygwin_dir   = r'C:\cygwin64\home\Aki\acoustic_model'
+cygwin_dir = r'C:\cygwin64\home\Aki\acoustic_model'
+kaldi_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5'
 #config_hcopy = os.path.join(cygwin_dir, 'config', 'config.HCopy')
 #config_train = os.path.join(cygwin_dir, 'config', 'config.train')
 config_hvite = os.path.join(cygwin_dir, 'config', 'config.HVite')

acoustic_model/performance_check.py

@@ -10,60 +10,56 @@ import re
 import numpy as np
 import pandas as pd
 import matplotlib.pyplot as plt
-#from sklearn.metrics import confusion_matrix
+from sklearn.metrics import confusion_matrix
 
 import acoustic_model_functions as am_func
 import convert_xsampa2ipa
 import defaultfiles as default
 
+from forced_alignment import pyhtk
+
 
 ## ======================= user define =======================
-#curr_dir   = r'C:\Users\Aki\source\repos\acoustic_model\acoustic_model'
-#config_ini = 'config.ini'
-#repo_dir   = r'C:\Users\Aki\source\repos'
-#forced_alignment_module     = repo_dir + '\\forced_alignment'
-#forced_alignment_module_old = repo_dir + '\\aki_tools'
-#ipa_xsampa_converter_dir    = repo_dir + '\\ipa-xsama-converter'
-#accent_classification_dir   = repo_dir + '\\accent_classification\accent_classification'
 excel_file = os.path.join(default.experiments_dir, 'stimmen', 'data', 'Frisian Variants Picture Task Stimmen.xlsx')
+data_dir   = os.path.join(default.experiments_dir, 'stimmen', 'data')
 
+wav_dir	= r'c:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\corpus\stimmen' # 16k
 
-#experiments_dir = r'C:\OneDrive\Research\rug\experiments'
-data_dir        = os.path.join(default.experiments_dir, 'stimmen', 'data')
-#csvfile         = data_dir + '\\Frisian Variants Picture Task Stimmen.csv'
-wav_dir	= os.path.join(default.experiments_dir, 'stimmen', 'wav')
 acoustic_model_dir = os.path.join(default.experiments_dir, 'friesian', 'acoustic_model', 'model')
-htk_dict_dir = os.path.join(default.experiments_dir, 'stimmen', 'dic_short')
+htk_dict_dir       = os.path.join(default.experiments_dir, 'stimmen', 'dic_short')
-fa_dir  = os.path.join(default.experiments_dir, 'stimmen', 'FA')
+fa_dir             = os.path.join(default.experiments_dir, 'stimmen', 'FA_44k')
+result_dir         = os.path.join(default.experiments_dir, 'stimmen', 'result')
 
-#cygwin_dir      = r'C:\cygwin64\home\Aki\acoustic_model'
+kaldi_data_dir = os.path.join(default.kaldi_dir, 'data', 'alignme') 
-#lex_asr		= os.path.join(default.fame_dir, 'lexicon', 'lex.asr')
+kaldi_dict_dir = os.path.join(default.kaldi_dir, 'data', 'local', 'dict')
+lexicon_txt    = os.path.join(kaldi_dict_dir, 'lexicon.txt')
+
+#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
-make_dic_files = 0
+make_htk_dict_files = 0
-do_forced_alignment_htk = 1
+do_forced_alignment_htk = 0
+eval_forced_alignment_htk = 0
 make_kaldi_data_files = 0
 make_kaldi_lexicon_txt = 0
-load_forced_alignment_kaldi = 0
+load_forced_alignment_kaldi = 1
-eval_forced_alignment = 0
+eval_forced_alignment_kaldi = 1
+
 
 
 
 ## ======================= add paths =======================
-
 sys.path.append(os.path.join(default.repo_dir, 'forced_alignment'))
 from forced_alignment import convert_phone_set
 from forced_alignment import pyhtk
 
 sys.path.append(os.path.join(default.repo_dir, 'toolbox'))
-#import pyHTK
 from evaluation import plot_confusion_matrix
 
 
 ## ======================= convert phones ======================
-
 mapping = convert_xsampa2ipa.load_converter('xsampa', 'ipa', default.ipa_xsampa_converter_dir)
 
 xls = pd.ExcelFile(excel_file)
@@ -115,11 +111,11 @@ df = pd.DataFrame({'filename': df['Filename'],
 # cleansing. 
 df = df[~df['famehtk'].isin(['/', ''])]
 
+word_list = np.unique(df['word'])
+
 
 ## ======================= make dict files used for HTK. ======================
-if make_dic_files:
+if make_htk_dict_files:
-    word_list = np.unique(df['word'])
-
     output_type = 3
     
     for word in word_list:
@@ -129,67 +125,67 @@ if make_dic_files:
         pronvar_ = df['famehtk'][df['word'].str.match(word)]
 
         # make dic file.
-        am_func.make_dic(word, pronvar_, htk_dict_file, output_type)
+        am_func.make_htk_dict(word, pronvar_, htk_dict_file, output_type)
         
 
 ## ======================= forced alignment using HTK =======================
 if do_forced_alignment_htk:
     
-    #hmm_num = 2
+    #for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]:
-    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 = []
+        predictions = pd.DataFrame({'filename': [''], 
+                    'word': [''], 
+                    'xsampa': [''],
+                    'ipa': [''],
+                    'famehtk': [''],
+                    'prediction': ['']})
         for i, filename in enumerate(df['filename']):
             print('=== {0}/{1} ==='.format(i, len(df)))
-            wav_file = os.path.join(wav_dir, 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):
+                    word = df['word'][i]
+                    WORD = word.upper()
+                    fa_file = os.path.join(fa_dir, filename.replace('.wav', '.txt') + hmm_num_str)
                     
-            if os.path.exists(wav_file) and i in df['filename'].keys():
+                    #if not os.path.exists(fa_file):
-                word = df['word'][i]
+                    # make label file.
-                WORD = word.upper()
+                    label_file = os.path.join(wav_dir, filename.replace('.wav', '.lab'))
+                    with open(label_file, 'w') as f:
+                        lines = f.write(WORD)
 
-                # make label file.
+                    htk_dict_file = os.path.join(htk_dict_dir, word + '.dic')
-                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, 
-                fa_file       = os.path.join(fa_dir, filename.replace('.wav', '.txt') + hmm_num_str)
+                                default.phonelist, acoustic_model)
-                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)
+                    prediction = am_func.read_fileFA(fa_file)
-                predictions.append(prediction)
 
-                os.remove(label_file)
+                    print('{0}: {1} -> {2}'.format(WORD, df['famehtk'][i], prediction))
-                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:
-                predictions.append('')
+                prediction = ''
-                print('!!!!! file not found.')
+                print('!!!!! invalid entry.')
 
-        predictions = np.array(predictions)
+        predictions.to_pickle(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.pkl'))
-        #match = np.c_[words[predictions != ''], pronunciations[predictions != ''], predictions[predictions != '']]
-        np.save(os.path.join(data_dir, 'predictions_hmm' + hmm_num_str + '.npy'), predictions)
 
 
 ## ======================= make files which is used for forced alignment by Kaldi =======================
 if make_kaldi_data_files:
-    wav_dir	 = r'c:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\corpus\stimmen'
-    kaldi_work_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5'
-    kaldi_data_dir = os.path.join(kaldi_work_dir, 'data', 'alignme') 
-    kaldi_dict_dir = os.path.join(kaldi_work_dir, 'data', 'local', 'dict')
-    htk_dict_dir   = os.path.join(experiments_dir, 'stimmen', 'dic_top3')
 
     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')
 
-    lexicon_txt = os.path.join(kaldi_dict_dir, 'lexicon.txt')
-
-    predictions = []
-    file_num_max = len(filenames)
-
     # remove previous files.
     if os.path.exists(wav_scp):
         os.remove(wav_scp)
@@ -203,12 +199,12 @@ if make_kaldi_data_files:
     f_utt2spk   = open(utt2spk, 'a', encoding="utf-8", newline='\n')
 
     # make wav.scp, text, and utt2spk files.
-    for i in range(0, file_num_max):
+    for i in df.index:
-    #for i in range(400, 410):
+        filename = df['filename'][i]
-        print('=== {0}/{1} ==='.format(i+1, file_num_max))
+        print('=== {0}: {1} ==='.format(i, filename))
-        filename = filenames[i]
-        wav_file = wav_dir + '\\' + 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', '')
@@ -222,7 +218,7 @@ if make_kaldi_data_files:
             f_wav_scp.write('{0} {1}\n'.format(utterance_id, wav_file_unix))
 
             # text file
-            word = words[i].lower()
+            word = df['word'][i].lower()
             f_text_file.write('{0}\t{1}\n'.format(utterance_id, word))
 
             # utt2spk
@@ -235,203 +231,257 @@ if make_kaldi_data_files:
 
 ## ======================= make lexicon txt which is used by Kaldi =======================
 if make_kaldi_lexicon_txt:
-    kaldi_work_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5'
+    option_num = 6
-    kaldi_dict_dir = os.path.join(kaldi_work_dir, 'data', 'local', 'dict')
-    lexicon_txt = os.path.join(kaldi_dict_dir, 'lexicon.txt')
-    option_num = 5
 
     # remove previous file.
     if os.path.exists(lexicon_txt):
         os.remove(lexicon_txt)
-
+    lexiconp_txt = lexicon_txt.replace('lexicon.txt', 'lexiconp.txt')
-    mapping = convert_xsampa2ipa.load_converter('xsampa', 'ipa', ipa_xsampa_converter_dir)
+    if os.path.exists(lexiconp_txt):
-    with open(csvfile, encoding="utf-8") as fin:
+        os.remove(lexiconp_txt)
-        lines = csv.reader(fin, delimiter=';', lineterminator="\n", skipinitialspace=True)
-        next(lines, None) # skip the headers
-
-        filenames	   = []
-        words		   = []
-        pronunciations = []
-        p = []
-        for line in lines:
-            if line[1] is not '' and len(line) > 5:
-                filenames.append(line[0])
-                words.append(line[1])
-                pron_xsampa = line[3]
-                pron_ipa = convert_xsampa2ipa.conversion('xsampa', 'ipa', mapping, pron_xsampa)
-                pron_ipa = pron_ipa.replace('ː', ':')
-            
-                # adjust to phones used in the acoustic model.
-                pronunciations.append(pron_ipa)
-
-    # check if all phones are in the phonelist of the acoustic model.
-    #'y', 'b', 'ɾ', 'u', 'ɔ:', 'ø', 't', 'œ', 'n', 'ɒ', 'ɐ', 'f', 'o', 'k', 'x', 'ɡ', 'v', 's', 'ɛ:', 'ɪ:', 'ɑ', 'ɛ', 'a', 'd', 'z', 'ɪ', 'ɔ', 'l', 'i:', 'm', 'p', 'a:', 'i', 'e', 'j', 'o:', 'ʁ', 'h', ':', 'e:', 'ə', 'æ', 'χ', 'w', 'r', 'ə:', 'sp', 'ʊ', 'u:', 'ŋ'
-
-    filenames	   = np.array(filenames)
-    words		   = np.array(words)
-    wordlist       = np.unique(words)
-    pronunciations = np.array(pronunciations)
                 
     # output lexicon.txt
-    #f_lexicon_txt = open(lexicon_txt, 'a', encoding="utf-8", newline='\n')
+    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.
-        pronvar_ = pronunciations[words == word]
+        pronunciation_variants = df['ipa'][df['word'].str.match(word)]
-        # remove ''
-        pronvar_ = np.delete(pronvar_, np.where(pronvar_==''))
 
-        c = Counter(pronvar_)
+        c = Counter(pronunciation_variants)
         total_num = sum(c.values())
 
-        for key, value in c.most_common(option_num):
+        #with open(result_dir + '\\' + word + '.csv', 'a', encoding="utf-8", newline='\n') as f:
-            #print('{0}\t{1}\t{2}\t{3}'.format(word, key, value, total_num))
+        #    for key in c.keys():
-            key = key.replace('æ', 'ɛ')
+        #        f.write("{0},{1}\n".format(key,c[key]))
-            key = key.replace('ɐ', 'a')
-            key = key.replace('ɑ', 'a')
-            key = key.replace('ɾ', 'r')
-            key = key.replace('ʁ', 'r')
-            key = key.replace('ʊ', 'u')
-            key = key.replace('χ', 'x')
-            #print('-->{0}\t{1}\t{2}\t{3}\n'.format(word, key, value, total_num))
 
+        for key, value in c.most_common(option_num):
             # make possible pronounciation variant list.
-            pronvar_list = [key]
+            pronvar_list = am_func.fame_pronunciation_variant(key)
-            while 'ø:' in ' '.join(pronvar_list) or 'œ' in ' '.join(pronvar_list) or 'ɒ' in ' '.join(pronvar_list):
-                pronvar_list_ = []
-                for p in pronvar_list:
-                    if 'ø:' in p:
-                        pronvar_list_.append(p.replace('ø:', 'ö'))
-                        pronvar_list_.append(p.replace('ø:', 'ö:'))
-                    if 'œ' in p:
-                        pronvar_list_.append(p.replace('œ', 'ɔ̈'))
-                        pronvar_list_.append(p.replace('œ', 'ɔ̈:'))
-                    if 'ɒ' in p:
-                        pronvar_list_.append(p.replace('ɒ', 'ɔ̈'))
-                        pronvar_list_.append(p.replace('ɒ', 'ɔ̈:'))
-                pronvar_list = np.unique(pronvar_list_)
 
             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])
 
-    # output
     pronvar_list_all = np.array(pronvar_list_all)
     pronvar_list_all = np.unique(pronvar_list_all, axis=0)
-    #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()
+    
+    # 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()
+
 
 ## ======================= load kaldi forced alignment result =======================
 if load_forced_alignment_kaldi:
-    kaldi_work_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5'
+    phones_txt = os.path.join(default.kaldi_dir, 'data', 'lang', 'phones.txt')
-    phones_txt = kaldi_work_dir + '\\data\\lang\\phones.txt'
+    merged_alignment_txt = os.path.join(default.kaldi_dir, 'exp', 'tri1_alignme', 'merged_alignment.txt')
-    merged_alignment_txt = kaldi_work_dir + '\\exp\\tri1_alignme\\merged_alignment.txt'
     
-    filenames	   = np.load(data_dir + '\\filenames.npy')
+    #filenames	   = np.load(data_dir + '\\filenames.npy')
-    words		   = np.load(data_dir + '\\words.npy')
+    #words		   = np.load(data_dir + '\\words.npy')
-    pronunciations = np.load(data_dir + '\\pronunciations_ipa.npy')
+    #pronunciations = np.load(data_dir + '\\pronunciations_ipa.npy')
-    pronvar_list_all = np.load(data_dir + '\\pronvar_list_all.npy')
+    #pronvar_list_all = np.load(data_dir + '\\pronvar_list_all.npy')
-    word_list = np.unique(words)    
+    #word_list = np.unique(words)    
 
     # load the mapping between phones and ids.
     with open(phones_txt, 'r', encoding="utf-8") as f:
-        mappings = f.read().split('\n')
+        mapping_phone2id = f.read().split('\n')
 
     phones = []
-    phone_ids = []
+    phone_ids = [] # ID of phones
-    for m in mappings:
+    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')
 
-    fa_filenames = []
+    predictions = pd.DataFrame({'filename': [''], 
-    fa_pronunciations = []
+                'word': [''], 
-    filename_ = ''
+                'xsampa': [''],
-    pron = []
+                'ipa': [''],
+                'famehtk': [''],
+                'prediction': ['']})
+    #fa_filenames = []
+    #fa_pronunciations = []
+    utterance_id_ = ''
+    pronunciation = []
     for line in lines:
         line = line.split(' ')
         if len(line) == 5:
-            filename = line[0]
+            utterance_id = line[0]
-            if filename == filename_:
+            if utterance_id == utterance_id_:
                 phone_id = int(line[4])
                 #if not phone_id == 1:
-                phone    = phones[phone_ids.index(phone_id)]
+                phone_ = phones[phone_ids.index(phone_id)]
-                pron_ = re.sub(r'_[A-Z]', '', phone)
+                phone  = re.sub(r'_[A-Z]', '', phone_)
-                if not pron_ == 'SIL':
+                if not phone == 'SIL':
-                    pron.append(pron_)
+                    pronunciation.append(phone)
             else:
-                fa_filenames.append(re.sub(r'speaker_[0-9]{4}-', '', filename))
+                filename = re.sub(r'speaker_[0-9]{4}-', '', utterance_id_)
-                fa_pronunciations.append(' '.join(pron))
+                prediction = ''.join(pronunciation)
-                pron = []
+                df_ = df[df['filename'].str.match(filename)]
-
+                df_idx = df_.index[0]
-            filename_ = filename
+                prediction_ = pd.Series([#filename,
-
+                                    #df_['word'][df_idx], 
-    # correct or not.
+                                    #df_['xsampa'][df_idx], 
-    #for filename, fa_pronunciation in zip(fa_filenames, fa_pronunciations):
+                                    #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:
+if eval_forced_alignment_htk:
-    match_num = []
+    htk_dict_dir = os.path.join(default.experiments_dir, 'stimmen', 'dic_short')
-    for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128, 256]:
-    #hmm_num = 256
-        hmm_num_str = str(hmm_num)
-        match = np.load(data_dir + '\\match_hmm' + hmm_num_str + '.npy')
 
-        # use dic_short?
+    compare_hmm_num = 1
-        if 1:
+
-            pronunciation_variants = np.array(['WORD', 'pronunciation']).reshape(1, 2)
+    if compare_hmm_num:
-            for word in word_list:
+        f_result = open(os.path.join(result_dir, 'result.csv'), 'w')
-                fileDic = experiments_dir + r'\stimmen\dic_top3' + '\\' + word + '.dic'
+        f_result.write("nmix,Oog,Oog,Oor,Oor,Pauw,Pauw,Reus,Reus,Reuzenrad,Reuzenrad,Roeiboot,Roeiboot,Rozen,Rozen\n")
-                pronunciation_variants = np.r_[pronunciation_variants, pyHTK.loadHTKdic(fileDic)]
+
+    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'))
+
+
+        # 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.
-            match_short = []
+            result_ = result[result['word'].str.match(word)]
-            for line in match:
+            result_ = result_[result_['famehtk'].isin(pronunciation_variants)]
-                word = line[0]
-                WORD = word.upper()
-                pronvar = pronunciation_variants[pronunciation_variants[:, 0] == word.upper(), 1]
 
-                if line[1] in pronvar:
+            match_num = sum(result_['famehtk'] == result_['prediction'])
-                    match_short.append(line)
+            total_num = len(result_)
 
-            match_short = np.array(match_short)
+            print("word '{0}': {1}/{2} ({3:.2f} %)".format(word, match_num, total_num, match_num/total_num*100))
-            match = np.copy(match_short)
+            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'])
 
-            # number of match
+                plt.figure()
-            total_match = sum(match[:, 1] == match[:, 2])
+                plot_confusion_matrix(cm, classes=pronunciation_variants, normalize=False)
-            print("{}: {}/{}".format(hmm_num_str, total_match, match.shape[0]))
+                plt.savefig(result_dir + '\\cm_' + word + '.png')
-            match_num.append([hmm_num, total_match, match.shape[0]])
+
+        if compare_hmm_num:
+            f_result.write('\n')
+
+    if compare_hmm_num:
+        f_result.close()
 
 
-    # number of mixtures vs accuracy
+## ======================= evaluate the result of forced alignment of kaldi =======================
-    match_num = np.array(match_num)
+if eval_forced_alignment_kaldi:
-    plt.xscale("log")
+    result = pd.read_pickle(os.path.join(result_dir, 'kaldi', 'predictions.pkl'))
-    plt.plot(match_num[:, 0], match_num[:, 1]/match_num[0, 2], 'o-')
-    plt.xlabel('number of mixtures', fontsize=14, fontweight='bold')
-    plt.ylabel('accuracy', fontsize=14, fontweight='bold')
-    plt.show()
 
-    # confusion matrix
+    f_result = open(os.path.join(result_dir, 'result.csv'), 'w')
-    #dir_out   = r'C:\OneDrive\Research\rug\experiments\stimmen\result'
+    f_result.write("word,total,valid,match,[%]\n")
-    #word_list = np.unique(match[:, 0])
 
-    #for word in word_list:
+    # load pronunciation variants
-    #    match_ = match[match[:, 0] == word, :]
+    with open(lexicon_txt, 'r', encoding="utf-8", newline='\n') as f:
-    #    cm = confusion_matrix(match_[:, 1], match_[:, 2])
+        lines = f.read().split('\n')[:-1]
-    #    pronvar = pronunciation_variants[pronunciation_variants[:, 0] == word.upper(), 1]
+        pronunciation_variants_all = [line.split('\t') for line in lines]
 
-    #    plt.figure()
+    word_list = np.delete(word_list, [0], 0) # remove 'Oog'
-    #    plot_confusion_matrix(cm, classes=pronvar, normalize=True)
+    for word in word_list:
-    #    plt.savefig(dir_out + '\\cm_' + word + '.png')
+
+        # 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')