if args.language == 'en':
from nemo_text_processing.text_normalization.en.clean_eval_data import filter_loaded_data
file_path = args.input
normalizer = Normalizer(input_case=args.input_case, lang=args.language)
print("Loading training data: " + file_path)
training_data = load_files([file_path])
if args.filter:
training_data = filter_loaded_data(training_data)
if args.category is None:
print("Sentence level evaluation...")
sentences_un_normalized, sentences_normalized, _ = training_data_to_sentences(training_data)
print("- Data: " + str(len(sentences_normalized)) + " sentences")
sentences_prediction = normalizer.normalize_list(sentences_un_normalized)
print("- Normalized. Evaluating...")
sentences_accuracy = evaluate(
preds=sentences_prediction, labels=sentences_normalized, input=sentences_un_normalized
)
print("- Accuracy: " + str(sentences_accuracy))
print("Token level evaluation...")
tokens_per_type = training_data_to_tokens(training_data, category=args.category)
token_accuracy = {}
for token_type in tokens_per_type:
print("- Token type: " + token_type)
tokens_un_normalized, tokens_normalized = tokens_per_type[token_type]
print(" - Data: " + str(len(tokens_normalized)) + " tokens")
tokens_prediction = normalizer.normalize_list(tokens_un_normalized)
print(" - Denormalized. Evaluating...")
"""
with open(file_path, 'w') as fp:
for line in data:
fp.write(line + '\n')
def parse_args():
parser = ArgumentParser()
parser.add_argument("--input", help="input file path", required=True, type=str)
parser.add_argument("--language", help="language", choices=['en'], default="en", type=str)
parser.add_argument("--output", help="output file path", required=True, type=str)
parser.add_argument(
"--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str
)
parser.add_argument("--verbose", help="print meta info for debugging", action='store_true')
return parser.parse_args()
if __name__ == "__main__":
args = parse_args()
file_path = args.input
normalizer = Normalizer(input_case=args.input_case, lang=args.language)
print("Loading data: " + file_path)
data = load_file(file_path)
print("- Data: " + str(len(data)) + " sentences")
normalizer_prediction = normalizer.normalize_list(data, verbose=args.verbose)
write_file(args.output, normalizer_prediction)
print(f"- Normalized. Writing out to {args.output}")
def split_text(
in_file: str,
out_file: str,
vocabulary: List[str] = None,
language='eng',
remove_brackets=True,
do_lower_case=True,
min_length=0,
max_length=100,
additional_split_symbols=None,
use_nemo_normalization=False,
):
"""
Breaks down the in_file roughly into sentences. Each sentence will be on a separate line.
Written form of the numbers will be converted to its spoken equivalent, OOV punctuation will be removed.
Args:
in_file: path to original transcript
out_file: path to the output file
vocabulary: ASR model vocabulary
language: text language
remove_brackets: Set to True if square [] and curly {} brackets should be removed from text.
Text in square/curly brackets often contains inaudible fragments like notes or translations
do_lower_case: flag that determines whether to apply lower case to the in_file text
min_length: Min number of chars of the text segment for alignment. Short segments will be combined to be
at least min_length (not recommended for multi speaker data).
max_length: Max number of chars of the text segment for alignment
additional_split_symbols: Additional symbols to use for sentence split if eos sentence split resulted in
segments longer than --max_length
use_nemo_normalization: Set to True to use NeMo normalization tool to convert numbers from written to spoken
format. Normalization using num2words will be applied afterwards to make sure there are no numbers present
in the text, otherwise they will be replaced with a space and that could deteriorate segmentation results.
"""
print(f'Splitting text in {in_file} into sentences.')
with open(in_file, "r") as f:
transcript = f.read()
# remove some symbols for better split into sentences
transcript = (transcript.replace("\n", " ").replace("\t", " ").replace(
"…",
"...").replace("\\",
" ").replace("--",
" -- ").replace(". . .",
"...").replace("‘", "’"))
# remove extra space
transcript = re.sub(r' +', ' ', transcript)
transcript = re.sub(r'(\.+)', '. ', transcript)
if remove_brackets:
transcript = re.sub(r'(\[.*?\])', ' ', transcript)
# remove text in curly brackets
transcript = re.sub(r'(\{.*?\})', ' ', transcript)
lower_case_unicode = ''
upper_case_unicode = ''
if language == 'ru':
lower_case_unicode = '\u0430-\u04FF'
upper_case_unicode = '\u0410-\u042F'
elif language not in ['ru', 'eng']:
print(
f'Consider using {language} unicode letters for better sentence split.'
)
# remove space in the middle of the lower case abbreviation to avoid splitting into separate sentences
matches = re.findall(
r'[a-z' + lower_case_unicode + ']\.\s[a-z' + lower_case_unicode +
']\.', transcript)
for match in matches:
transcript = transcript.replace(match, match.replace('. ', '.'))
# find phrases in quotes
with_quotes = re.finditer(r'“[A-Za-z ?]+.*?”', transcript)
sentences = []
last_idx = 0
for m in with_quotes:
match = m.group()
match_idx = m.start()
if last_idx < match_idx:
sentences.append(transcript[last_idx:match_idx])
sentences.append(match)
last_idx = m.end()
sentences.append(transcript[last_idx:])
sentences = [s.strip() for s in sentences if s.strip()]
# Read and split transcript by utterance (roughly, sentences)
split_pattern = f"(?<!\w\.\w.)(?<![A-Z{upper_case_unicode}][a-z{lower_case_unicode}]\.)(?<![A-Z{upper_case_unicode}]\.)(?<=\.|\?|\!|\.”|\?”\!”)\s"
new_sentences = []
for sent in sentences:
new_sentences.extend(regex.split(split_pattern, sent))
sentences = [s.strip() for s in new_sentences if s.strip()]
def additional_split(sentences, split_on_symbols, max_length):
if len(split_on_symbols) == 0:
return sentences
split_on_symbols = split_on_symbols.split('|')
def _split(sentences, delimiter, max_length):
result = []
for s in sentences:
if len(s) <= max_length:
result.append(s)
else:
split_sent = s.split(delimiter)
result.extend([s + delimiter for s in split_sent[:-1]] +
[split_sent[-1]])
return result
another_sent_split = []
for sent in sentences:
split_sent = [sent]
for delimiter in split_on_symbols:
split_sent = _split(split_sent, delimiter + ' ', max_length)
another_sent_split.extend(split_sent)
sentences = [s.strip() for s in another_sent_split if s.strip()]
return sentences
sentences = additional_split(sentences, additional_split_symbols,
max_length)
# check to make sure there will be no utterances for segmentation with only OOV symbols
vocab_no_space_with_digits = set(vocabulary + [i for i in range(10)])
vocab_no_space_with_digits.remove(' ')
sentences = [
s for s in sentences
if len(vocab_no_space_with_digits.intersection(set(s))) > 0
]
if min_length > 0:
sentences_comb = []
sentences_comb.append(sentences[0])
# combines short sentence
for i in range(1, len(sentences)):
if len(sentences_comb[-1]) < min_length or len(
sentences[i]) < min_length:
sentences_comb[-1] += ' ' + sentences[i].strip()
else:
sentences_comb.append(sentences[i].strip())
sentences = sentences_comb
sentences = [s.strip() for s in sentences if s.strip()]
# save split text with original punctuation and case
out_dir, out_file_name = os.path.split(out_file)
with open(os.path.join(out_dir, out_file_name[:-4] + '_with_punct.txt'),
"w") as f:
f.write("\n".join(sentences))
# substitute common abbreviations before applying lower case
if language == 'ru':
for k, v in RU_ABBREVIATIONS.items():
sentences = [s.replace(k, v) for s in sentences]
if language == 'ru':
# replace Latin characters with Russian
for k, v in LATIN_TO_RU.items():
sentences = [s.replace(k, v) for s in sentences]
if language == 'eng' and use_nemo_normalization:
if not NEMO_NORMALIZATION_AVAILABLE:
raise ValueError(f'NeMo normalization tool is not installed.')
print('Using NeMo normalization tool...')
normalizer = Normalizer(input_case='cased')
sentences_norm = normalizer.normalize_list(sentences, verbose=False)
if len(sentences_norm) != len(sentences):
raise ValueError(
f'Normalization failed, number of sentences does not match.')
sentences = '\n'.join(sentences)
# replace numbers with num2words
try:
p = re.compile("\d+")
new_text = ''
match_end = 0
for i, m in enumerate(p.finditer(sentences)):
match = m.group()
match_start = m.start()
if i == 0:
new_text = sentences[:match_start]
else:
new_text += sentences[match_end:match_start]
match_end = m.end()
new_text += sentences[match_start:match_end].replace(
match, num2words(match, lang=language))
new_text += sentences[match_end:]
sentences = new_text
except NotImplementedError:
print(
f'{language} might be missing in "num2words" package. Add required language to the choices for the'
f'--language argument.')
raise
sentences = (sentences.replace("’", "'").replace("»", '"').replace(
"«",
'"').replace("\\", "").replace("”", '"').replace("„", '"').replace(
"´", "'").replace("-- --", "--").replace("--", " -- ").replace(
"’", "'").replace('“', '"').replace('“', '"').replace(
"‘", "'").replace('—', '-').replace("- -", "--").replace(
'`',
"'").replace(' !', '!').replace(' ?', '?').replace(
' ,', ',').replace(' .', '.').replace(
' ;', ';').replace(' :', ':').replace(
'!!', '!').replace('--', '-').replace(
'“',
'"').replace(', , ',
', ').replace('=', ''))
allowed_punct = [',', '.', '?', '!', ':', ';', '-', '"', '(', ')']
# clean up normalized text and keep only allowed_punct and ASR vocabulary (lower and upper case)
symbols_to_remove = ''.join(
set(sentences).difference(
set(vocabulary + [s.upper()
for s in vocabulary] + ['\n'] + allowed_punct)))
sentences_norm = sentences.translate(''.maketrans(
symbols_to_remove,
len(symbols_to_remove) * ' '))
with open(
os.path.join(out_dir,
out_file_name[:-4] + '_with_punct_normalized.txt'),
"w") as f:
f.write(sentences_norm)
if do_lower_case:
sentences = sentences.lower()
# remove all OOV symbols
symbols_to_remove = ''.join(
set(sentences).difference(set(vocabulary + ['\n'])))
sentences = sentences.translate(''.maketrans(symbols_to_remove,
len(symbols_to_remove) * ' '))
# remove extra space
sentences = re.sub(r' +', ' ', sentences)
with open(out_file, "w") as f:
f.write(sentences)
def split_text(
in_file: str,
out_file: str,
vocabulary: List[str],
language="en",
remove_brackets=True,
do_lower_case=True,
max_length=100,
additional_split_symbols=None,
use_nemo_normalization=False,
):
"""
Breaks down the in_file roughly into sentences. Each sentence will be on a separate line.
Written form of the numbers will be converted to its spoken equivalent, OOV punctuation will be removed.
Args:
in_file: path to original transcript
out_file: path to the output file
vocabulary: ASR model vocabulary
language: text language
remove_brackets: Set to True if square [] and curly {} brackets should be removed from text.
Text in square/curly brackets often contains inaudible fragments like notes or translations
do_lower_case: flag that determines whether to apply lower case to the in_file text
max_length: Max number of words of the text segment for alignment
additional_split_symbols: Additional symbols to use for sentence split if eos sentence split resulted in
segments longer than --max_length
use_nemo_normalization: Set to True to use NeMo normalization tool to convert numbers from written to spoken
format. Normalization using num2words will be applied afterwards to make sure there are no numbers present
in the text, otherwise they will be replaced with a space and that could deteriorate segmentation results.
"""
print(f"Splitting text in {in_file} into sentences.")
with open(in_file, "r") as f:
transcript = f.read()
# remove some symbols for better split into sentences
transcript = (transcript.replace("\n", " ").replace("\t", " ").replace(
"…", "...").replace("\\", " ").replace("--",
" -- ").replace(". . .", "..."))
# remove extra space
transcript = re.sub(r" +", " ", transcript)
transcript = re.sub(r"(\.+)", ". ", transcript)
if remove_brackets:
transcript = re.sub(r'(\[.*?\])', ' ', transcript)
# remove text in curly brackets
transcript = re.sub(r'(\{.*?\})', ' ', transcript)
lower_case_unicode = ''
upper_case_unicode = ''
if language == "ru":
lower_case_unicode = '\u0430-\u04FF'
upper_case_unicode = '\u0410-\u042F'
elif language not in ["ru", "en"]:
print(
f"Consider using {language} unicode letters for better sentence split."
)
# remove space in the middle of the lower case abbreviation to avoid splitting into separate sentences
matches = re.findall(
r'[a-z' + lower_case_unicode + ']\.\s[a-z' + lower_case_unicode +
']\.', transcript)
for match in matches:
transcript = transcript.replace(match, match.replace('. ', '.'))
# find phrases in quotes
with_quotes = re.finditer(r'“[A-Za-z ?]+.*?”', transcript)
sentences = []
last_idx = 0
for m in with_quotes:
match = m.group()
match_idx = m.start()
if last_idx < match_idx:
sentences.append(transcript[last_idx:match_idx])
sentences.append(match)
last_idx = m.end()
sentences.append(transcript[last_idx:])
sentences = [s.strip() for s in sentences if s.strip()]
# Read and split transcript by utterance (roughly, sentences)
split_pattern = f"(?<!\w\.\w.)(?<![A-Z{upper_case_unicode}][a-z{lower_case_unicode}]\.)(?<![A-Z{upper_case_unicode}]\.)(?<=\.|\?|\!|\.”|\?”\!”)\s"
new_sentences = []
for sent in sentences:
new_sentences.extend(regex.split(split_pattern, sent))
sentences = [s.strip() for s in new_sentences if s.strip()]
def additional_split(sentences, split_on_symbols):
if len(split_on_symbols) == 0:
return sentences
split_on_symbols = split_on_symbols.split("|")
def _split(sentences, delimiter):
result = []
for sent in sentences:
split_sent = sent.split(delimiter)
# keep the delimiter
split_sent = [(s + delimiter).strip()
for s in split_sent[:-1]] + [split_sent[-1]]
if "," in delimiter:
# split based on comma usually results in too short utterance, combine sentences
# that result in a single word split. It's usually not recommended to do that for other delimiters.
comb = []
for s in split_sent:
MIN_LEN = 2
# if the previous sentence is too short, combine it with the current sentence
if len(comb) > 0 and (len(comb[-1].split()) <= MIN_LEN
or len(s.split()) <= MIN_LEN):
comb[-1] = comb[-1] + " " + s
else:
comb.append(s)
result.extend(comb)
else:
result.extend(split_sent)
return result
another_sent_split = []
for sent in sentences:
split_sent = [sent]
for delimiter in split_on_symbols:
split_sent = _split(split_sent, delimiter + " ")
another_sent_split.extend(split_sent)
sentences = [s.strip() for s in another_sent_split if s.strip()]
return sentences
sentences = additional_split(sentences, additional_split_symbols)
vocabulary_symbols = []
for x in vocabulary:
if x != "<unk>":
# for BPE models
vocabulary_symbols.extend(
[x for x in x.replace("##", "").replace("▁", "")])
vocabulary_symbols = list(set(vocabulary_symbols))
vocabulary_symbols += [x.upper() for x in vocabulary_symbols]
# check to make sure there will be no utterances for segmentation with only OOV symbols
vocab_no_space_with_digits = set(vocabulary_symbols +
[str(i) for i in range(10)])
if " " in vocab_no_space_with_digits:
vocab_no_space_with_digits.remove(" ")
sentences = [
s.strip() for s in sentences
if len(vocab_no_space_with_digits.intersection(set(s.lower()))) > 0
and s.strip()
]
# when no punctuation marks present in the input text, split based on max_length
if len(sentences) == 1:
sent = sentences[0].split()
sentences = []
for i in range(0, len(sent), max_length):
sentences.append(" ".join(sent[i:i + max_length]))
sentences = [s.strip() for s in sentences if s.strip()]
# save split text with original punctuation and case
out_dir, out_file_name = os.path.split(out_file)
with open(os.path.join(out_dir, out_file_name[:-4] + "_with_punct.txt"),
"w") as f:
f.write(re.sub(r' +', ' ', "\n".join(sentences)))
# substitute common abbreviations before applying lower case
if language == "ru":
for k, v in RU_ABBREVIATIONS.items():
sentences = [s.replace(k, v) for s in sentences]
# replace Latin characters with Russian
for k, v in LATIN_TO_RU.items():
sentences = [s.replace(k, v) for s in sentences]
if language == "en" and use_nemo_normalization:
if not NEMO_NORMALIZATION_AVAILABLE:
raise ValueError("NeMo normalization tool is not installed.")
print("Using NeMo normalization tool...")
normalizer = Normalizer(input_case="cased",
cache_dir=os.path.join(
os.path.dirname(out_file), "en_grammars"))
sentences_norm = normalizer.normalize_list(sentences,
verbose=False,
punct_post_process=True)
if len(sentences_norm) != len(sentences):
raise ValueError(
"Normalization failed, number of sentences does not match.")
else:
sentences = sentences_norm
sentences = '\n'.join(sentences)
# replace numbers with num2words
try:
p = re.compile("\d+")
new_text = ""
match_end = 0
for i, m in enumerate(p.finditer(sentences)):
match = m.group()
match_start = m.start()
if i == 0:
new_text = sentences[:match_start]
else:
new_text += sentences[match_end:match_start]
match_end = m.end()
new_text += sentences[match_start:match_end].replace(
match, num2words(match, lang=language))
new_text += sentences[match_end:]
sentences = new_text
except NotImplementedError:
print(
f"{language} might be missing in 'num2words' package. Add required language to the choices for the"
f"--language argument.")
raise
sentences = re.sub(r' +', ' ', sentences)
with open(
os.path.join(out_dir,
out_file_name[:-4] + "_with_punct_normalized.txt"),
"w") as f:
f.write(sentences)
if do_lower_case:
sentences = sentences.lower()
symbols_to_remove = ''.join(
set(sentences).difference(set(vocabulary_symbols + ["\n", " "])))
sentences = sentences.translate(''.maketrans(symbols_to_remove,
len(symbols_to_remove) * " "))
# remove extra space
sentences = re.sub(r' +', ' ', sentences)
with open(out_file, "w") as f:
f.write(sentences)
