import os
import audiomate
from audiomate import annotations
from audiomate import issuers
from audiomate.utils import textfile
from . import base
from . import downloader
DOWNLOAD_URL = 'http://www.openslr.org/resources/17/musan.tar.gz'
AUDIO_TYPES_ = ['music', 'noise', 'speech']
ANN_NUM_COLUMS_ = {'music': 4, 'noise': 1, 'speech': 3}
ANN_FILE_NAME_ = 'ANNOTATIONS'
[docs]class MusanDownloader(downloader.ArchiveDownloader):
"""
Downloader for the MUSAN Corpus.
Args:
url (str): The url to download the dataset from. If not given the default URL is used.
It is expected to be a tar.gz file.
"""
def __init__(self, url=None):
if url is None:
url = DOWNLOAD_URL
super(MusanDownloader, self).__init__(
url,
move_files_up=True
)
[docs] @classmethod
def type(cls):
return 'musan'
[docs]class MusanReader(base.CorpusReader):
"""
Reader for the MUSAN corpus. MUSAN is a corpus of music, speech, and noise recordings.
.. seealso::
`MUSAN: A Music, Speech, and Noise Corpus <https://arxiv.org/pdf/1510.08484v1.pdf>`_
Paper explaining the structure and characteristics of the corpus
`OpenSLR: MUSAN <http://www.openslr.org/17/>`_
Download page
"""
[docs] @classmethod
def type(cls):
return 'musan'
def _check_for_missing_files(self, path):
# Some label files are missing anyway in the original data set
# (e.g. speech/us-gov/ANNOTATIONS). What's left would be checking for missing directories.
return []
def _load(self, path):
create_or_get_issuer = {
'music': self._create_or_get_music_issuer,
'noise': self._create_or_get_noise_issuer,
'speech': self._create_or_get_speech_issuer,
}
corpus = audiomate.Corpus(path=path)
for type_name, type_directory in self._directories(path).items():
for _, source_directory in self._directories(type_directory).items():
labels_path = os.path.join(source_directory, ANN_FILE_NAME_)
labels = {}
if os.path.exists(labels_path):
labels = textfile.read_separated_lines_with_first_key(
labels_path, separator=' ', max_columns=ANN_NUM_COLUMS_[type_name])
it = os.scandir(source_directory)
for entry in it:
if not entry.name.endswith('.wav'):
continue
file_path = os.path.join(source_directory, entry.name)
file_idx = entry.name[0:-4] # chop of .wav
utterance_idx = file_idx # every file is a separate utterance
issuer_idx = create_or_get_issuer[type_name](corpus, file_idx, labels)
corpus.new_file(file_path, track_idx=file_idx, copy_file=False)
utterance = corpus.new_utterance(utterance_idx, file_idx, issuer_idx)
utterance.set_label_list(annotations.LabelList.create_single(
type_name, idx=audiomate.corpus.LL_DOMAIN))
return corpus
@staticmethod
def _directories(path):
directories = {}
it = os.scandir(path)
for entry in it:
if not entry.is_dir():
continue
directories[entry.name] = os.path.join(path, entry.name)
return directories
# noinspection PyUnusedLocal
@staticmethod
def _create_or_get_noise_issuer(corpus, file_idx, labels):
return None
@staticmethod
def _create_or_get_music_issuer(corpus, file_idx, labels):
if file_idx not in labels:
return None
issuer_idx = labels[file_idx][2]
if issuer_idx not in corpus.issuers:
issuer = issuers.Artist(issuer_idx, name=issuer_idx)
corpus.import_issuers(issuer)
return issuer_idx
@staticmethod
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