p280427/acoustic_model
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

The Stimmen excel file is loaded as Data Frame. Default values are given by defaultfiles.py.

Browse Source
This commit is contained in:
yemaozi88 2018-09-15 17:28:57 +02:00
parent 3a98e184fe
commit ea30b5c503
1 changed files with 232 additions and 143 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

375
acoustic_model/performance_check.py
View File

@ -10,7 +10,7 @@ import re
import numpy as np import numpy as np
import pandas as pd import pandas as pd
import matplotlib.pyplot as plt 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 acoustic_model_functions as am_func
import convert_xsampa2ipa import convert_xsampa2ipa
@ -31,22 +31,31 @@ excel_file = os.path.join(default.experiments_dir, 'stimmen', 'data', 'Frisian V
#experiments_dir = r'C:\OneDrive\Research\rug\experiments' #experiments_dir = r'C:\OneDrive\Research\rug\experiments'
data_dir = os.path.join(default.experiments_dir, 'stimmen', 'data') data_dir = os.path.join(default.experiments_dir, 'stimmen', 'data')
#csvfile = data_dir + '\\Frisian Variants Picture Task Stimmen.csv' #csvfile = data_dir + '\\Frisian Variants Picture Task Stimmen.csv'
wav_dir = os.path.join(default.experiments_dir, 'stimmen', 'wav') #wav_dir = os.path.join(default.experiments_dir, 'stimmen', 'wav_44k') # 44.1k
wav_dir = r'c:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\corpus\stimmen' # 16k
#wav_dir = r'c:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\corpus\stimmen' # 16k
acoustic_model_dir = os.path.join(default.experiments_dir, 'friesian', 'acoustic_model', 'model') 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')
kaldi_data_dir = os.path.join(default.kaldi_dir, 'data', 'alignme')
kaldi_dict_dir = os.path.join(default.kaldi_dir, 'data', 'local', 'dict')
lexicon_txt = os.path.join(kaldi_dict_dir, 'lexicon.txt')
#cygwin_dir = r'C:\cygwin64\home\Aki\acoustic_model' #cygwin_dir = r'C:\cygwin64\home\Aki\acoustic_model'
#lex_asr = os.path.join(default.fame_dir, 'lexicon', 'lex.asr') #lex_asr = os.path.join(default.fame_dir, 'lexicon', 'lex.asr')
#lex_asr_htk = os.path.join(default.fame_dir, 'lexicon', 'lex.asr_htk') #lex_asr_htk = os.path.join(default.fame_dir, 'lexicon', 'lex.asr_htk')
from forced_alignment import pyhtk
# procedure # procedure
make_dic_files = 0 make_dic_files = 0
do_forced_alignment_htk = 1 do_forced_alignment_htk = 0
make_kaldi_data_files = 0 make_kaldi_data_files = 0
make_kaldi_lexicon_txt = 0 make_kaldi_lexicon_txt = 0
load_forced_alignment_kaldi = 0 load_forced_alignment_kaldi = 1
eval_forced_alignment = 0 eval_forced_alignment = 0
@ -115,11 +124,11 @@ df = pd.DataFrame({'filename': df['Filename'],
# cleansing. # cleansing.
df = df[~df['famehtk'].isin(['/', ''])] df = df[~df['famehtk'].isin(['/', ''])]
word_list = np.unique(df['word'])
## ======================= make dict files used for HTK. ====================== ## ======================= make dict files used for HTK. ======================
if make_dic_files: if make_dic_files:
word_list = np.unique(df['word'])
output_type = 3 output_type = 3
for word in word_list: for word in word_list:
@ -134,61 +143,73 @@ if make_dic_files:
## ======================= forced alignment using HTK ======================= ## ======================= forced alignment using HTK =======================
if do_forced_alignment_htk: if do_forced_alignment_htk:
#hmm_num = 2 #hmm_num = 2
for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]: #for hmm_num in [1]:
#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) hmm_num_str = str(hmm_num)
acoustic_model = os.path.join(acoustic_model_dir, 'hmm' + hmm_num_str + r'-2\hmmdefs') 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']): for i, filename in enumerate(df['filename']):
print('=== {0}/{1} ==='.format(i, len(df))) 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) and i in df['filename'].keys(): if os.path.exists(wav_file):
word = df['word'][i] word = df['word'][i]
WORD = word.upper() 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)
# 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)
predictions.append(prediction)
os.remove(label_file) prediction = am_func.read_fileFA(fa_file)
print('{0}: {1} -> {2}'.format(WORD, df['famehtk'][i], prediction)) #predictions.append(prediction)
print('{0}: {1} -> {2}'.format(WORD, df['famehtk'][i], prediction))
else:
prediction = ''
#predictions.append('')
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: else:
predictions.append('') prediction = ''
print('!!!!! file not found.') #predictions.append('')
print('!!!!! invalid entry.')
predictions = np.array(predictions)
#match = np.c_[words[predictions != ''], pronunciations[predictions != ''], predictions[predictions != '']] #predictions = np.array(predictions)
np.save(os.path.join(data_dir, 'predictions_hmm' + hmm_num_str + '.npy'), predictions) #np.save(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.npy'), predictions)
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 ======================= ## ======================= make files which is used for forced alignment by Kaldi =======================
if make_kaldi_data_files: 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') wav_scp = os.path.join(kaldi_data_dir, 'wav.scp')
text_file = os.path.join(kaldi_data_dir, 'text') text_file = os.path.join(kaldi_data_dir, 'text')
utt2spk = os.path.join(kaldi_data_dir, 'utt2spk') utt2spk = os.path.join(kaldi_data_dir, 'utt2spk')
lexicon_txt = os.path.join(kaldi_dict_dir, 'lexicon.txt') #predictions = []
#file_num_max = len(filenames)
predictions = []
file_num_max = len(filenames)
# remove previous files. # remove previous files.
if os.path.exists(wav_scp): if os.path.exists(wav_scp):
@ -203,30 +224,42 @@ if make_kaldi_data_files:
f_utt2spk = open(utt2spk, 'a', encoding="utf-8", newline='\n') f_utt2spk = open(utt2spk, 'a', encoding="utf-8", newline='\n')
# make wav.scp, text, and utt2spk files. # make wav.scp, text, and utt2spk files.
for i in range(0, file_num_max): predictions = pd.DataFrame({'filename': [''],
'word': [''],
'xsampa': [''],
'ipa': [''],
'famehtk': [''],
'prediction': ['']})
#for i in range(0, file_num_max):
#for i in range(400, 410): #for i in range(400, 410):
print('=== {0}/{1} ==='.format(i+1, file_num_max)) for i, filename in enumerate(df['filename']):
filename = filenames[i]
#print('=== {0}/{1} ==='.format(i+1, file_num_max))
#filename = filenames[i]
print('=== {0}/{1} ==='.format(i, len(df)))
wav_file = wav_dir + '\\' + filename 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', '')
utterance_id = utterance_id.replace(' ', '_')
utterance_id = speaker_id + '-' + utterance_id
if os.path.exists(wav_file): # wav.scp file
speaker_id = 'speaker_' + str(i).zfill(4) wav_file_unix = wav_file.replace('\\', '/')
utterance_id = filename.replace('.wav', '') wav_file_unix = wav_file_unix.replace('c:/', '/mnt/c/')
utterance_id = utterance_id.replace(' ', '_')
utterance_id = speaker_id + '-' + utterance_id
# wav.scp file f_wav_scp.write('{0} {1}\n'.format(utterance_id, wav_file_unix))
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 = words[i].lower()
word = df['word'][i].lower()
f_text_file.write('{0}\t{1}\n'.format(utterance_id, word))
# text file # utt2spk
word = words[i].lower() f_utt2spk.write('{0} {1}\n'.format(utterance_id, speaker_id))
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_wav_scp.close()
f_text_file.close() f_text_file.close()
@ -235,54 +268,58 @@ if make_kaldi_data_files:
## ======================= make lexicon txt which is used by Kaldi ======================= ## ======================= make lexicon txt which is used by Kaldi =======================
if make_kaldi_lexicon_txt: if make_kaldi_lexicon_txt:
kaldi_work_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5' #lexicon_txt = os.path.join(kaldi_dict_dir, 'lexicon.txt')
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 option_num = 5
# remove previous file. # remove previous file.
if os.path.exists(lexicon_txt): if os.path.exists(lexicon_txt):
os.remove(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)
mapping = convert_xsampa2ipa.load_converter('xsampa', 'ipa', ipa_xsampa_converter_dir) #mapping = convert_xsampa2ipa.load_converter('xsampa', 'ipa', ipa_xsampa_converter_dir)
with open(csvfile, encoding="utf-8") as fin:
lines = csv.reader(fin, delimiter=';', lineterminator="\n", skipinitialspace=True)
next(lines, None) # skip the headers
filenames = [] #with open(csvfile, encoding="utf-8") as fin:
words = [] # lines = csv.reader(fin, delimiter=';', lineterminator="\n", skipinitialspace=True)
pronunciations = [] # next(lines, None) # skip the headers
p = []
for line in lines: # filenames = []
if line[1] is not '' and len(line) > 5: # words = []
filenames.append(line[0]) # pronunciations = []
words.append(line[1]) # p = []
pron_xsampa = line[3] # for line in lines:
pron_ipa = convert_xsampa2ipa.conversion('xsampa', 'ipa', mapping, pron_xsampa) # if line[1] is not '' and len(line) > 5:
pron_ipa = pron_ipa.replace('ː', ':') # 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. # # adjust to phones used in the acoustic model.
pronunciations.append(pron_ipa) # pronunciations.append(pron_ipa)
# check if all phones are in the phonelist of the acoustic model. ## 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' ##'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) #filenames = np.array(filenames)
words = np.array(words) #words = np.array(words)
wordlist = np.unique(words) #wordlist = np.unique(words)
pronunciations = np.array(pronunciations) #pronunciations = np.array(pronunciations)
# output lexicon.txt # 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 = [] pronvar_list_all = []
for word in word_list: for word in word_list:
# pronunciation variant of the target word. # pronunciation variant of the target word.
pronvar_ = pronunciations[words == word] #pronvar_ = pronunciations[words == word]
pronunciation_variants = df['ipa'][df['word'].str.match(word)]
#pronunciation_variants = np.unique(pronunciation_variants)
# remove '' # remove ''
pronvar_ = np.delete(pronvar_, np.where(pronvar_=='')) #pronvar_ = np.delete(pronvar_, np.where(pronvar_==''))
c = Counter(pronvar_) c = Counter(pronunciation_variants)
total_num = sum(c.values()) total_num = sum(c.values())
for key, value in c.most_common(option_num): for key, value in c.most_common(option_num):
@ -291,7 +328,9 @@ if make_kaldi_lexicon_txt:
key = key.replace('ɐ', 'a') key = key.replace('ɐ', 'a')
key = key.replace('ɑ', 'a') key = key.replace('ɑ', 'a')
key = key.replace('ɾ', 'r') key = key.replace('ɾ', 'r')
key = key.replace('ɹ', 'r') # ???
key = key.replace('ʁ', 'r') key = key.replace('ʁ', 'r')
key = key.replace('ʀ', 'r') # ???
key = key.replace('ʊ', 'u') key = key.replace('ʊ', 'u')
key = key.replace('χ', 'x') key = key.replace('χ', 'x')
#print('-->{0}\t{1}\t{2}\t{3}\n'.format(word, key, value, total_num)) #print('-->{0}\t{1}\t{2}\t{3}\n'.format(word, key, value, total_num))
@ -320,23 +359,24 @@ if make_kaldi_lexicon_txt:
# output # output
pronvar_list_all = np.array(pronvar_list_all) pronvar_list_all = np.array(pronvar_list_all)
pronvar_list_all = np.unique(pronvar_list_all, axis=0) pronvar_list_all = np.unique(pronvar_list_all, axis=0)
#f_lexicon_txt.write('<UNK>\tSPN\n') f_lexicon_txt.write('<UNK>\tSPN\n')
#for line in pronvar_list_all: for line in pronvar_list_all:
# f_lexicon_txt.write('{0}\t{1}\n'.format(line[0].lower(), line[1])) 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 ======================= ## ======================= load kaldi forced alignment result =======================
if load_forced_alignment_kaldi: if load_forced_alignment_kaldi:
kaldi_work_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5' kaldi_work_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5'
phones_txt = kaldi_work_dir + '\\data\\lang\\phones.txt' phones_txt = os.path.join(kaldi_work_dir, 'data', 'lang', 'phones.txt')
merged_alignment_txt = kaldi_work_dir + '\\exp\\tri1_alignme\\merged_alignment.txt' merged_alignment_txt = os.path.join(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. # load the mapping between phones and ids.
with open(phones_txt, 'r', encoding="utf-8") as f: with open(phones_txt, 'r', encoding="utf-8") as f:
@ -379,59 +419,108 @@ if load_forced_alignment_kaldi:
# correct or not. # correct or not.
#for filename, fa_pronunciation in zip(fa_filenames, fa_pronunciations): #for filename, fa_pronunciation in zip(fa_filenames, fa_pronunciations):
# 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)
# #predictions.append(prediction)
# print('{0}: {1} -> {2}'.format(WORD, df['famehtk'][i], prediction))
# else:
# prediction = ''
# #predictions.append('')
# 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 = ''
# #predictions.append('')
# print('!!!!! invalid entry.')
# #predictions = np.array(predictions)
# #np.save(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.npy'), predictions)
# predictions.to_pickle(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.pkl'))
## ======================= evaluate the result of forced alignment ======================= ## ======================= evaluate the result of forced alignment =======================
if eval_forced_alignment: if eval_forced_alignment:
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 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")
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) hmm_num_str = str(hmm_num)
match = np.load(data_dir + '\\match_hmm' + hmm_num_str + '.npy') if compare_hmm_num:
f_result.write("{},".format(hmm_num_str))
# use dic_short?
if 1: #match = np.load(data_dir + '\\match_hmm' + hmm_num_str + '.npy')
pronunciation_variants = np.array(['WORD', 'pronunciation']).reshape(1, 2) #prediction = np.load(os.path.join(result_dir, 'htk', 'predictions_hmm' + hmm_num_str + '.npy'))
for word in word_list: #prediction = pd.Series(prediction, index=df.index, name='prediction')
fileDic = experiments_dir + r'\stimmen\dic_top3' + '\\' + word + '.dic' #result = pd.concat([df, prediction], axis=1)
pronunciation_variants = np.r_[pronunciation_variants, pyHTK.loadHTKdic(fileDic)] 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. # 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_short.append(line)
match_short = np.array(match_short) match_num = sum(result_['famehtk'] == result_['prediction'])
match = np.copy(match_short) total_num = len(result_)
# number of match print("word '{0}': {1}/{2} ({3:.2f} %)".format(word, match_num, total_num, match_num/total_num*100))
total_match = sum(match[:, 1] == match[:, 2]) if compare_hmm_num:
print("{}: {}/{}".format(hmm_num_str, total_match, match.shape[0])) f_result.write("{0},{1},".format(match_num, total_num))
match_num.append([hmm_num, total_match, match.shape[0]]) 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')
# number of mixtures vs accuracy if compare_hmm_num:
match_num = np.array(match_num) f_result.write('\n')
plt.xscale("log")
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 if compare_hmm_num:
#dir_out = r'C:\OneDrive\Research\rug\experiments\stimmen\result' f_result.close()
#word_list = np.unique(match[:, 0])
#for word in word_list:
# match_ = match[match[:, 0] == word, :]
# cm = confusion_matrix(match_[:, 1], match_[:, 2])
# pronvar = pronunciation_variants[pronunciation_variants[:, 0] == word.upper(), 1]
# plt.figure()
# plot_confusion_matrix(cm, classes=pronvar, normalize=True)
# plt.savefig(dir_out + '\\cm_' + word + '.png')