def tatoeba_loader(keep_archive):
"""
Download, extract and convert the Tatoeba archive.
Then build all possible CSV files (e.g. `<dataset_name>_train.csv`, `<dataset_name>_test.csv`).
Args:
keep_archive (bool): Keep or delete the downloaded archive afterwards.
Returns:
List[str]: List containing the created CSV file paths.
"""
# Download and extract the dataset if necessary.
download.maybe_download(__URL, md5=__MD5, cache_archive=keep_archive)
if not os.path.isdir(__SOURCE_PATH):
raise ValueError('"{}" is not a directory.'.format(__SOURCE_PATH))
# Download user ratings CSV file.
csv_path = os.path.join(__SOURCE_PATH, 'users_sentences.csv')
download.download_with_progress(
'http://downloads.tatoeba.org/exports/users_sentences.csv', csv_path)
assert os.path.exists(csv_path)
target = 'train'
# Generate the WAV and a string for the `<target>.txt` file.
output = __tatoeba_loader(target)
# Generate the `<target>.txt` file.
csv_path = generate_csv(__NAME, target, output)
# Cleanup extracted folder.
download.cleanup_cache(__FOLDER_NAME)
return csv_path
def tedlium_loader(keep_archive):
"""
Download, extract and convert the TEDLIUM archive.
Then build all possible CSV files (e.g. `<dataset_name>_train.csv`, `<dataset_name>_test.csv`).
Requires lots of disk space, since the original format (SPH) is converted to WAV and then split
up into parts.
Args:
keep_archive (bool): Keep or delete the downloaded archive afterwards.
Returns:
List[str]: List containing the created CSV file paths.
"""
# Download and extract the dataset if necessary.
download.maybe_download(__URL, md5=__MD5, cache_archive=keep_archive)
if not os.path.isdir(__SOURCE_PATH):
raise ValueError('"{}" is not a directory.'.format(__SOURCE_PATH))
# Folders for each target.
targets = [{
'name': 'train',
'folder': 'train'
}, {
'name': 'test',
'folder': 'test'
}, {
'name': 'dev',
'folder': 'dev'
}]
txt_paths = []
for target in targets:
# Create target folder if necessary.
target_directory = os.path.join(__TARGET_PATH, target['folder'], 'sph')
if not os.path.exists(target_directory):
os.makedirs(target_directory)
# Generate the WAV and a string for the `<target>.txt` file.
source_directory = os.path.join(__SOURCE_PATH, target['folder'])
output = __tedlium_loader(source_directory)
# Generate the `<target>.txt` file.
txt_paths.append(generate_csv(__NAME, target['name'], output))
# Cleanup extracted folder.
download.cleanup_cache(__FOLDER_NAME)
return tuple(txt_paths)
def common_voice_loader(keep_archive):
"""
Download, extract and convert the Common Voice archive.
Then build all possible CSV files (e.g. `<dataset_name>_train.csv`, `<dataset_name>_test.csv`).
Uses only the valid datasets, additional constraints are:
* Downvotes must be at maximum 1/4 of upvotes.
* Valid accents are: 'us', 'england', 'canada', 'australia'.
* Accepting samples with only 1 upvote at the moment.
Args:
keep_archive (bool): Keep or delete the downloaded archive afterwards.
Returns:
List[str]: List containing the created CSV file paths.
"""
# Download and extract the dataset if necessary.
download.maybe_download(__URL, md5=__MD5, cache_archive=keep_archive)
if not os.path.isdir(__SOURCE_PATH):
raise ValueError('"{}" is not a directory.'.format(__SOURCE_PATH))
# Folders for each target.
targets = [{
'name': 'train',
'folders': ['cv-valid-train']
}, {
'name': 'test',
'folders': ['cv-valid-test']
}, {
'name': 'dev',
'folders': ['cv-valid-dev']
}]
csv_paths = []
for target in targets:
# Generate the path and label for the `<target>.csv` file.
output = __common_voice_loader(target['folders'])
# Generate the `<target>.csv` file.
csv_paths.append(generate_csv(__NAME, target['name'], output))
# Cleanup extracted folder.
download.cleanup_cache(__FOLDER_NAME)
return csv_paths
def libri_speech_loader(keep_archive):
"""
Download, extract and convert the Libri Speech archive.
Then build all possible CSV files (e.g. `<dataset_name>_train.csv`, `<dataset_name>_test.csv`).
Args:
keep_archive (bool): Keep or delete the downloaded archive afterwards.
Returns:
List[str]: List containing the created CSV file paths.
"""
# Download and extract the dataset if necessary.
download.maybe_download_batch(__URLS,
md5s=__MD5S,
cache_archives=keep_archive)
if not os.path.isdir(__SOURCE_PATH):
raise ValueError('"{}" is not a directory.'.format(__SOURCE_PATH))
# Folders for each target.
targets = [{
'name': 'train',
'folders': ['train-clean-100', 'train-clean-360']
}, {
'name': 'test',
'folders': ['test-clean']
}, {
'name': 'dev',
'folders': ['dev-clean']
}]
csv_paths = []
for target in targets:
# Generate the WAV and a string for the `<target>.txt` file.
output = __libri_speech_loader(target['folders'])
# Generate the `<target>.txt` file.
csv_paths.append(generate_csv(__NAME, target['name'], output))
# Cleanup extracted folder.
download.cleanup_cache(__FOLDER_NAME)
return csv_paths