acoustic_model/acoustic_model/performance_check.py

import os
import sys
import csv
import subprocess
import configparser
from collections import Counter

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt


## ======================= functions =======================

def read_fileFA(fileFA):
	"""
	read the result file of HTK forced alignment.
	this function only works when input is one word.
	"""
	with open(fileFA, 'r') as f:
		lines = f.read()
		lines = lines.split('\n')

	phones = []
	for line in lines:
		line_split = line.split()
		if len(line_split) > 1:
			phones.append(line_split[2])

	return ' '.join(phones)


#####################
##   USER DEFINE   ##
#####################
curr_dir   = r'C:\Users\Aki\source\repos\acoustic_model\acoustic_model'
config_ini = curr_dir + '\\config.ini'
forced_alignment_module     = r'C:\Users\Aki\source\repos\forced_alignment'
forced_alignment_module_old = r'C:\OneDrive\Research\rug\code\forced_alignment\forced_alignment'
ipa_xsampa_converter_dir = r'C:\Users\Aki\source\repos\ipa-xsama-converter'

csvfile = r"C:\OneDrive\Research\rug\stimmen\Frisian Variants Picture Task Stimmen.csv"
experiments_dir = r'C:\OneDrive\Research\rug\experiments'
data_dir = experiments_dir + '\\stimmen\\data'
cygwin_dir = r'C:\cygwin64\home\Aki\acoustic_model'

# procedure
convert_phones = 0
make_dic_files = 0
make_dic_files_short = 0
do_forced_alignment = 0
eval_forced_alignment = 1



## ======================= add paths =======================

sys.path.append(forced_alignment_module)
from forced_alignment import convert_phone_set

# for interactive window
sys.path.append(curr_dir)
import convert_xsampa2ipa
import acoustic_model_functions as am_func

# for forced-alignment
sys.path.append(forced_alignment_module_old)
import pyHTK


## ======================= load variables =======================
config = configparser.ConfigParser()
config.sections()
config.read(config_ini)

FAME_dir = config['Settings']['FAME_dir']

lex_asr		= FAME_dir + '\\lexicon\\lex.asr'
lex_asr_htk = FAME_dir + '\\lexicon\\lex.asr_htk'


## ======================= convert phones ======================
if convert_phones:
	mapping = convert_xsampa2ipa.load_converter('xsampa', 'ipa', ipa_xsampa_converter_dir)

	## check phones included in FAME!
	# the phones used in the lexicon.
	#phonelist = am_func.get_phonelist(lex_htk)

	# the lines which include a specific phone.
	#lines = am_func.find_phone(lex_asr, 'x')

	with open(csvfile, encoding="utf-8") as fin:
		lines = csv.reader(fin, delimiter=';', lineterminator="\n", skipinitialspace=True)
		next(lines, None) # skip the headers

		filenames	   = []
		words		   = []
		pronunciations = []
		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', ':')
				pron_famehtk = convert_phone_set.ipa2famehtk(pron_ipa)
			
				# adjust to phones used in the acoustic model.
				pron_famehtk = pron_famehtk.replace('sp', 'sil')
				pron_famehtk = pron_famehtk.replace('ce :', 'ce') # because ceh is ignored.
				pron_famehtk = pron_famehtk.replace('w :', 'wh')
				pron_famehtk = pron_famehtk.replace('e :', 'eh')
				pron_famehtk = pron_famehtk.replace('eh :', 'eh')
				pron_famehtk = pron_famehtk.replace('ih :', 'ih')

				#translation_key = {'sp': 'sil', 'ce :': 'ceh', 'w :': 'wh'}
				#pron = []
				#for phoneme in pron_famehtk.split(' '):
				#	pron.append(translation_key.get(phoneme, phoneme))
				#pronunciations.append(' '.join(pron_famehtk))
				pronunciations.append(pron_famehtk)

	# check if all phones are in the phonelist of the acoustic model.
	#phonelist = ' '.join(pronunciations)
	#np.unique(phonelist.split(' '))
	#phonelist.find(':')

	filenames	   = np.array(filenames)
	words		   = np.array(words)
	pronunciations = np.array(pronunciations)

	del line, lines
	del pron_xsampa, pron_ipa, pron_famehtk

	np.save(data_dir + '\\filenames.npy', filenames)
	np.save(data_dir + '\\words.npy', words)
	np.save(data_dir + '\\pronunciations.npy', pronunciations)
else:
	filenames	   = np.load(data_dir + '\\filenames.npy')
	words		   = np.load(data_dir + '\\words.npy')
	
	pronunciations = np.load(data_dir + '\\pronunciations.npy')
word_list = np.unique(words)


## ======================= make dict files used for HTK. ======================
if make_dic_files:
	output_dir = experiments_dir + r'\stimmen\dic'

	for word in word_list:
		WORD	= word.upper()
		fileDic = output_dir + '\\' + word + '.dic'

		# make dic file.
		pronvar_ = pronunciations[words == word]
		pronvar  = np.unique(pronvar_)

		with open(fileDic, 'w') as f:
			for pvar in pronvar:
				f.write('{0}\t{1}\n'.format(WORD, pvar))


## ======================= make dict files for most popular words. ======================	
if make_dic_files_short:
	output_dir = experiments_dir + r'\stimmen\dic'

	#word = word_list[3]
	for word in word_list:
		WORD	= word.upper()
		fileStat = output_dir + '\\' + word + '_stat.csv'
		
		pronvar = pronunciations[words == word]
		c = Counter(pronvar)
		total_num = sum(c.values())

		with open(fileStat, 'w') as f:
			for key, value in c.items():
				f.write('{0}\t{1:.2f}\t{2}\t{3}\n'.format(value, value/total_num*100, WORD, key))


## ======================= forced alignment =======================
if do_forced_alignment:
	configHVite   = cygwin_dir + r'\config\config.HVite'
	filePhoneList = experiments_dir + r'\friesian\acoustic_model\config\phonelist_friesian.txt'
	wav_dir	= experiments_dir + r'\stimmen\wav'

	#for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128]:
	for hmm_num in [64]:
		hmm_num_str = str(hmm_num)
		AcousticModel = experiments_dir + r'\friesian\acoustic_model\model\hmm' + hmm_num_str + r'-3\hmmdefs'

		predictions = []
		file_num_max = len(filenames)
		for i in range(0, file_num_max):
			print('=== {0}/{1} ==='.format(i, file_num_max))
			filename = filenames[i]
			fileWav = wav_dir + '\\' + filename
	
			if os.path.exists(fileWav):
				word = words[i]
				WORD = word.upper()

				# make label file.
				fileLab = wav_dir + '\\' + filename.replace('.wav', '.lab') 
				with open(fileLab, 'w') as f:
					lines = f.write(WORD)

				fileDic = experiments_dir + r'\stimmen\dic_short' + '\\' + word + '.dic'
				fileFA  = experiments_dir + r'\stimmen\FA_short' + '\\' + filename.replace('.wav', '.txt') + hmm_num_str

				pyHTK.doHVite(fileWav, fileLab, fileDic, fileFA, configHVite, filePhoneList, AcousticModel)
				prediction = read_fileFA(fileFA)
				predictions.append(prediction)

				os.remove(fileLab)
				print('{0}: {1} -> {2}'.format(WORD, pronunciations[i], prediction))
			else:
				predictions.append('')
				print('!!!!! file not found.')

		predictions = np.array(predictions)
		match = np.c_[words[predictions != ''], pronunciations[predictions != ''], predictions[predictions != '']]
		np.save(data_dir + '\\match_hmm' + hmm_num_str + '.npy', match)


## ======================= evaluate the result of forced alignment =======================
if eval_forced_alignment:

	#for hmm_num in [1, 2, 4, 8, 16, 32, 64]:
	hmm_num = 64
	hmm_num_str = str(hmm_num)
	match = np.load(data_dir + '\\match_hmm' + hmm_num_str + '.npy')
	
	# use dic_short?
	if 1:
		pronunciation_variants = np.array(['WORD', 'pronunciation']).reshape(1, 2)
		for word in word_list:
			fileDic = experiments_dir + r'\stimmen\dic_short' + '\\' + word + '.dic'
			pronunciation_variants = np.r_[pronunciation_variants, pyHTK.loadHTKdic(fileDic)]

		match_short = []
		for line in match:
			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 = np.copy(match_short)

	# number of match
	total_match = sum(match[:, 1] == match[:, 2])
	print("{}: {}/{}".format(hmm_num_str, total_match, match.shape[0]))