def read_genders(genders_path, corpus):
if os.path.isfile(genders_path):
speakers = textfile.read_key_value_lines(genders_path,
separator=' ')
for speaker_idx, gender_str in speakers.items():
if gender_str == 'm':
gender = issuers.Gender.MALE
else:
gender = issuers.Gender.FEMALE
speaker = issuers.Speaker(speaker_idx, gender=gender)
corpus.import_issuers(speaker)
def load_books_of_speaker(self, corpus, path, speaker):
"""
Load all utterances for the speaker at the given path.
"""
utt_ids = []
for book_path in MailabsReader.get_folders(path):
meta_path = os.path.join(book_path, 'metadata.csv')
wavs_path = os.path.join(book_path, 'wavs')
meta = textfile.read_separated_lines(meta_path,
separator='|',
max_columns=3)
for entry in meta:
file_basename = entry[0]
transcription_raw = entry[1]
transcription_clean = entry[2]
if speaker is None:
idx = file_basename
utt_speaker = issuers.Speaker(idx)
speaker_idx = idx
corpus.import_issuers(utt_speaker)
else:
idx = '{}-{}'.format(speaker.idx, file_basename)
speaker_idx = speaker.idx
wav_name = '{}.wav'.format(file_basename)
wav_path = os.path.join(wavs_path, wav_name)
if os.path.isfile(
wav_path) and idx not in self.invalid_utterance_ids:
corpus.new_file(wav_path, idx)
ll_raw = annotations.LabelList.create_single(
transcription_raw,
idx=audiomate.corpus.LL_WORD_TRANSCRIPT_RAW)
ll_clean = annotations.LabelList.create_single(
transcription_clean,
idx=audiomate.corpus.LL_WORD_TRANSCRIPT)
utterance = corpus.new_utterance(idx, idx, speaker_idx)
utterance.set_label_list(ll_raw)
utterance.set_label_list(ll_clean)
utt_ids.append(utterance.idx)
return utt_ids
def _create_or_get_speech_issuer(corpus, file_idx, labels):
if file_idx not in labels:
return None
issuer = issuers.Speaker(file_idx)
if file_idx in labels:
if labels[file_idx][0] == 'm':
issuer.gender = issuers.Gender.MALE
elif labels[file_idx][0] == 'f':
issuer.gender = issuers.Gender.FEMALE
corpus.import_issuers(issuer)
return file_idx
def parse_speaker_info(readme_path):
""" Parse speaker info and return tuple (idx, gender). """
idx = None
gender = issuers.Gender.UNKNOWN
age_group = issuers.AgeGroup.UNKNOWN
native_lang = None
with open(readme_path, 'r', errors='ignore') as f:
for raw_line in f:
line = raw_line.strip()
if line is not None and line != '':
line = line.rstrip(';.')
parts = line.split(':', maxsplit=1)
if len(parts) > 1:
key = parts[0].strip().lower()
value = parts[1].strip()
if key == 'user name':
idx = value
value = value.lower()
if key == 'gender':
if value in ['männlich', 'male', 'mnnlich']:
gender = issuers.Gender.MALE
elif value in ['weiblich', 'female', '[female]']:
gender = issuers.Gender.FEMALE
if key == 'age range':
if value in ['erwachsener', 'adult', '[adult]', '[erwachsener]']:
age_group = issuers.AgeGroup.ADULT
elif value in ['senior', '[senior']:
age_group = issuers.AgeGroup.SENIOR
elif value in ['youth', 'jugendlicher', '[youth]', '[jugendlicher]']:
age_group = issuers.AgeGroup.YOUTH
elif value in ['kind', 'child']:
age_group = issuers.AgeGroup.CHILD
if key == 'language':
if value in ['de', 'ger', 'deu', '[de]']:
native_lang = 'deu'
elif value in ['en', 'eng', '[en]']:
native_lang = 'eng'
return issuers.Speaker(idx, gender=gender, age_group=age_group, native_language=native_lang)
def load_speaker(corpus, path):
""" Create a speaker instance for the given path. """
base_path, speaker_name = os.path.split(path)
base_path, gender_desc = os.path.split(base_path)
base_path, _ = os.path.split(base_path)
base_path, _ = os.path.split(base_path)
gender = issuers.Gender.UNKNOWN
if gender_desc == 'male':
gender = issuers.Gender.MALE
elif gender_desc == 'female':
gender = issuers.Gender.FEMALE
speaker = issuers.Speaker(speaker_name, gender=gender)
corpus.import_issuers(speaker)
return speaker
def _load_folder(folder_entry, corpus):
""" Load the given subfolder into the corpus (e.g. bed, one, ...) """
for wav_path in glob.glob(os.path.join(folder_entry.path, '*.wav')):
wav_name = os.path.basename(wav_path)
basename, __ = os.path.splitext(wav_name)
command = folder_entry.name
file_idx = '{}_{}'.format(basename, command)
issuer_idx = str(basename).split('_', maxsplit=1)[0]
corpus.new_file(wav_path, file_idx)
if issuer_idx not in corpus.issuers.keys():
corpus.import_issuers(issuers.Speaker(issuer_idx))
utt = corpus.new_utterance(file_idx, file_idx, issuer_idx)
labels = annotations.LabelList.create_single(
command, idx=audiomate.corpus.LL_WORD_TRANSCRIPT)
utt.set_label_list(labels)
def _load(self, path):
corpus = audiomate.Corpus(path=path)
for file_path in glob.glob(os.path.join(path, 'recordings', '*.wav')):
file_idx = os.path.splitext(os.path.basename(file_path))[0]
corpus.new_file(file_path, file_idx)
idx_parts = file_idx.split('_')
digit = idx_parts[0]
issuer_idx = '_'.join(idx_parts[1:-1])
if issuer_idx not in corpus.issuers.keys():
issuer = issuers.Speaker(issuer_idx)
corpus.import_issuers(issuer)
utterance = corpus.new_utterance(file_idx, file_idx, issuer_idx)
utterance.set_label_list(
annotations.LabelList.create_single(
str(digit), idx=audiomate.corpus.LL_WORD_TRANSCRIPT))
return corpus
def create_assets_if_needed(corpus, path, entry):
""" Create File/Utterance/Issuer, if they not already exist and return utt-idx. """
file_name = entry[1]
file_idx, _ = os.path.splitext(file_name)
if file_idx in INVALID_UTTS:
return None
if file_idx not in corpus.utterances.keys():
speaker_idx = entry[0]
transcription = entry[2]
if len(entry) >= 6:
age = CommonVoiceReader.map_age(entry[5])
else:
age = issuers.AgeGroup.UNKNOWN
if len(entry) >= 7:
gender = CommonVoiceReader.map_gender(entry[6])
else:
gender = issuers.Gender.UNKNOWN
file_path = os.path.join(path, 'clips', file_name)
corpus.new_file(file_path, file_idx)
if speaker_idx in corpus.issuers.keys():
issuer = corpus.issuers[speaker_idx]
else:
issuer = issuers.Speaker(speaker_idx,
gender=gender,
age_group=age)
corpus.import_issuers(issuer)
utterance = corpus.new_utterance(file_idx, file_idx, issuer.idx)
utterance.set_label_list(
annotations.LabelList.create_single(
transcription, idx=audiomate.corpus.LL_WORD_TRANSCRIPT))
return file_idx
def _load(self, path):
corpus = audiomate.Corpus(path=path)
data_path = os.path.join(path, 'data')
meta_data = AudioMNISTReader.load_speaker_meta(path)
for speaker_idx in os.listdir(data_path):
speaker_path = os.path.join(data_path, speaker_idx)
if os.path.isdir(speaker_path):
for file_path in glob.glob(os.path.join(speaker_path,
'*.wav')):
file_idx = os.path.splitext(os.path.basename(file_path))[0]
corpus.new_file(file_path, file_idx)
idx_parts = file_idx.split('_')
digit = idx_parts[0]
if speaker_idx not in corpus.issuers.keys():
issuer = issuers.Speaker(
speaker_idx,
gender=AudioMNISTReader.get_gender(
meta_data, speaker_idx),
age_group=AudioMNISTReader.get_age_group(
meta_data, speaker_idx))
corpus.import_issuers(issuer)
utterance = corpus.new_utterance(file_idx, file_idx,
speaker_idx)
utterance.set_label_list(
annotations.LabelList.create_single(
str(digit),
idx=audiomate.corpus.LL_WORD_TRANSCRIPT))
return corpus
def generate_issuers(n, rand=None):
if rand is None:
rand = random.Random()
items = []
for issuer_index in range(n):
issuer_idx = 'issuer-{}'.format(issuer_index)
issuer_type = rand.randint(1, 3)
if issuer_type == 1:
issuer = issuers.Speaker(issuer_idx,
gender=issuers.Gender.UNKNOWN,
age_group=issuers.AgeGroup.CHILD,
native_language='de')
elif issuer_type == 2:
issuer = issuers.Artist(issuer_idx, 'badam')
else:
issuer = issuers.Issuer(issuer_idx)
items.append(issuer)
return items
def load_file(self, folder_path, idx, corpus):
"""
Load speaker, file, utterance, labels
for the file with the given id.
"""
xml_path = os.path.join(folder_path, '{}.xml'.format(idx))
wav_paths = []
for wav_suffix in WAV_FILE_SUFFIXES:
wav_path = os.path.join(folder_path,
'{}_{}.wav'.format(idx, wav_suffix))
wav_name = os.path.split(wav_path)[1]
wav_idx = os.path.splitext(wav_name)[0]
if os.path.isfile(
wav_path) and wav_idx not in self.invalid_utterance_ids:
wav_paths.append(wav_path)
if len(wav_paths) == 0:
return []
with open(xml_path, 'r', encoding='utf-8') as f:
text = f.read()
transcription = TudaReader.extract_value(text, TRANSCRIPTION_PATTERN,
'transcription', xml_path)
transcription_raw = TudaReader.extract_value(
text, RAW_TRANSCRIPTION_PATTERN, 'raw_transcription', xml_path)
gender = TudaReader.extract_value(text, GENDER_PATTERN, 'gender',
xml_path)
is_native = TudaReader.extract_value(text, NATIVE_PATTERN, 'native',
xml_path)
age_class = TudaReader.extract_value(text, AGE_PATTERN, 'age',
xml_path)
speaker_idx = TudaReader.extract_value(text, SPEAKER_IDX_PATTERN,
'speaker_idx', xml_path)
if speaker_idx not in corpus.issuers.keys():
start_age_class = int(age_class.split('-')[0])
if start_age_class < 12:
age_group = issuers.AgeGroup.CHILD
elif start_age_class < 18:
age_group = issuers.AgeGroup.YOUTH
elif start_age_class < 65:
age_group = issuers.AgeGroup.ADULT
else:
age_group = issuers.AgeGroup.SENIOR
native_lang = None
if is_native == 'Ja':
native_lang = 'deu'
issuer = issuers.Speaker(speaker_idx,
gender=issuers.Gender(gender),
age_group=age_group,
native_language=native_lang)
corpus.import_issuers(issuer)
utt_ids = []
for wav_path in wav_paths:
wav_name = os.path.split(wav_path)[1]
wav_idx = os.path.splitext(wav_name)[0]
corpus.new_file(wav_path, wav_idx)
utt = corpus.new_utterance(wav_idx, wav_idx, speaker_idx)
utt.set_label_list(
annotations.LabelList.create_single(
transcription, idx=audiomate.corpus.LL_WORD_TRANSCRIPT))
utt.set_label_list(
annotations.LabelList.create_single(
transcription_raw,
idx=audiomate.corpus.LL_WORD_TRANSCRIPT_RAW))
utt_ids.append(wav_idx)
return utt_ids
def load_file(folder_path, idx, corpus):
"""
Load speaker, file, utterance, labels for the file with the given id.
"""
xml_path = os.path.join(folder_path, '{}.xml'.format(idx))
wav_paths = glob.glob(os.path.join(folder_path,
'{}_*.wav'.format(idx)))
if len(wav_paths) == 0:
return []
xml_file = open(xml_path, 'r', encoding='utf-8')
soup = BeautifulSoup(xml_file, 'lxml')
transcription = soup.recording.cleaned_sentence.string
transcription_raw = soup.recording.sentence.string
gender = soup.recording.gender.string
is_native = soup.recording.muttersprachler.string
age_class = soup.recording.ageclass.string
speaker_idx = soup.recording.speaker_id.string
if speaker_idx not in corpus.issuers.keys():
start_age_class = int(age_class.split('-')[0])
if start_age_class < 12:
age_group = issuers.AgeGroup.CHILD
elif start_age_class < 18:
age_group = issuers.AgeGroup.YOUTH
elif start_age_class < 65:
age_group = issuers.AgeGroup.ADULT
else:
age_group = issuers.AgeGroup.SENIOR
native_lang = None
if is_native == 'Ja':
native_lang = 'deu'
issuer = issuers.Speaker(speaker_idx,
gender=issuers.Gender(gender),
age_group=age_group,
native_language=native_lang)
corpus.import_issuers(issuer)
utt_ids = []
for wav_path in wav_paths:
wav_name = os.path.split(wav_path)[1]
wav_idx = os.path.splitext(wav_name)[0]
corpus.new_file(wav_path, wav_idx)
utt = corpus.new_utterance(wav_idx, wav_idx, speaker_idx)
utt.set_label_list(
annotations.LabelList.create_single(
transcription, idx=audiomate.corpus.LL_WORD_TRANSCRIPT))
utt.set_label_list(
annotations.LabelList.create_single(
transcription_raw,
idx=audiomate.corpus.LL_WORD_TRANSCRIPT_RAW))
utt_ids.append(wav_idx)
return utt_ids
def _load(self, path):
corpus = audiomate.Corpus(path=path)
for part in ['TEST', 'TRAIN']:
part_path = os.path.join(path, part)
part_utt_ids = set()
for region in os.listdir(part_path):
region_path = os.path.join(part_path, region)
if os.path.isdir(region_path):
for speaker_abbr in os.listdir(region_path):
speaker_path = os.path.join(region_path, speaker_abbr)
speaker_idx = speaker_abbr[1:]
if speaker_idx not in corpus.issuers.keys():
issuer = issuers.Speaker(speaker_idx)
if speaker_abbr[:1] == 'M':
issuer.gender = issuers.Gender.MALE
elif speaker_abbr[:1] == 'F':
issuer.gender = issuers.Gender.FEMALE
corpus.import_issuers(issuer)
for wav_path in glob.glob(
os.path.join(speaker_path, '*.WAV')):
sentence_idx = os.path.splitext(
os.path.basename(wav_path))[0]
utt_idx = '{}-{}-{}'.format(
region, speaker_abbr, sentence_idx).lower()
part_utt_ids.add(utt_idx)
raw_text_path = os.path.join(
speaker_path, '{}.TXT'.format(sentence_idx))
raw_text = textfile.read_separated_lines(
raw_text_path, separator=' ',
max_columns=3)[0][2]
words_path = os.path.join(
speaker_path, '{}.WRD'.format(sentence_idx))
words = textfile.read_separated_lines(
words_path, separator=' ', max_columns=3)
phones_path = os.path.join(
speaker_path, '{}.PHN'.format(sentence_idx))
phones = textfile.read_separated_lines(
phones_path, separator=' ', max_columns=3)
corpus.new_file(wav_path, utt_idx)
utt = corpus.new_utterance(utt_idx, utt_idx,
speaker_idx)
raw_ll = annotations.LabelList.create_single(
raw_text,
idx=audiomate.corpus.LL_WORD_TRANSCRIPT_RAW)
utt.set_label_list(raw_ll)
word_ll = annotations.LabelList(
idx=audiomate.corpus.LL_WORD_TRANSCRIPT)
for record in words:
start = int(record[0]) / 16000
end = int(record[1]) / 16000
word_ll.addl(record[2], start=start, end=end)
utt.set_label_list(word_ll)
phone_ll = annotations.LabelList(
idx=audiomate.corpus.LL_PHONE_TRANSCRIPT)
for record in phones:
start = int(record[0]) / 16000
end = int(record[1]) / 16000
phone_ll.addl(record[2], start=start, end=end)
utt.set_label_list(phone_ll)
utt_filter = subset.MatchingUtteranceIdxFilter(
utterance_idxs=part_utt_ids)
subview = subset.Subview(corpus, filter_criteria=[utt_filter])
corpus.import_subview(part, subview)
return corpus
def _load(self, path):
corpus = audiomate.Corpus()
article_paths = sorted(self.get_articles(path))
reader_map = {}
file_map = {}
for article_idx, article_path in enumerate(article_paths):
audio_files = self.get_audio_file_info(article_path)
reader_name, reader_gender = self.get_reader_info(article_path)
segments = self.get_segments(article_path)
if reader_name not in reader_map.keys():
speaker = issuers.Speaker(
'{:0>8}'.format(len(reader_map)),
gender=reader_gender
)
reader_map[reader_name] = speaker
corpus.import_issuers(speaker)
else:
speaker = reader_map[reader_name]
for start, end, text in segments:
file_path = self.find_audio_file_for_segment(start, end, audio_files)
if file_path is not None:
if file_path not in file_map.keys():
track = tracks.FileTrack(
'{:0>10}'.format(len(file_map)),
file_path
)
file_map[file_path] = track
corpus.import_tracks(track)
else:
track = file_map[file_path]
track_offset = audio_files[file_path]
utt_start = start - track_offset
utt_end = end - track_offset
utt_idx = '{}_{}_{}_{}'.format(
speaker.idx,
track.idx,
int(start * 1000),
int(end * 1000)
)
if utt_idx not in self.invalid_utterance_ids:
utt = corpus.new_utterance(
utt_idx,
track.idx,
issuer_idx=speaker.idx,
start=utt_start,
end=utt_end
)
ll = annotations.LabelList.create_single(
text,
audiomate.corpus.LL_WORD_TRANSCRIPT
)
utt.set_label_list(ll)
return audiomate.Corpus.from_corpus(corpus)
