def __init__(
self,
manifest_path=None,
min_snr_db=10,
max_snr_db=50,
max_gain_db=300.0,
rng=None,
audio_tar_filepaths=None,
shuffle_n=100,
orig_sr=16000,
):
self._manifest = collections.ASRAudioText(manifest_path,
parser=parsers.make_parser(
[]),
index_by_file_id=True)
self._audiodataset = None
self._tarred_audio = False
self._orig_sr = orig_sr
self._data_iterator = None
if audio_tar_filepaths:
self._tarred_audio = True
self._audiodataset = AugmentationDataset(manifest_path,
audio_tar_filepaths,
shuffle_n)
self._data_iterator = iter(self._audiodataset)
self._rng = random.Random() if rng is None else rng
self._min_snr_db = min_snr_db
self._max_snr_db = max_snr_db
self._max_gain_db = max_gain_db
def __init__(self, manifest_path: str, tar_filepaths: Union[str, List[str]], shuffle_n: int = 128):
self._manifest = collections.ASRAudioText(manifest_path, parser=parsers.make_parser([]), index_by_file_id=True)
if isinstance(tar_filepaths, str):
# Replace '(' and '[' with '{'
brace_keys_open = ['(', '[', '<', '_OP_']
for bkey in brace_keys_open:
if bkey in tar_filepaths:
tar_filepaths = tar_filepaths.replace(bkey, "{")
# Replace ')' and ']' with '}'
brace_keys_close = [')', ']', '>', '_CL_']
for bkey in brace_keys_close:
if bkey in tar_filepaths:
tar_filepaths = tar_filepaths.replace(bkey, "}")
if not HAVE_OMEGACONG_WEBDATASET:
raise LightningNotInstalledException(self)
self.audio_dataset = wd.WebDataset(urls=tar_filepaths, nodesplitter=None)
if shuffle_n > 0:
self.audio_dataset = self.audio_dataset.shuffle(shuffle_n)
else:
logging.info("WebDataset will not shuffle files within the tar files.")
self.audio_dataset = self.audio_dataset.rename(audio='wav', key='__key__').to_tuple('audio', 'key')
self.audio_iter = iter(self.audio_dataset)
def __init__(
self,
manifest_filepath: str,
parser: Union[str, Callable],
max_duration: Optional[float] = None,
min_duration: Optional[float] = None,
max_utts: int = 0,
bos_id: Optional[int] = None,
eos_id: Optional[int] = None,
pad_id: int = 0,
index_by_file_id: bool = False,
):
self.parser = parser
self.collection = collections.ASRAudioText(
manifests_files=manifest_filepath,
parser=parser,
min_duration=min_duration,
max_duration=max_duration,
max_number=max_utts,
index_by_file_id=index_by_file_id,
)
self.eos_id = eos_id
self.bos_id = bos_id
self.pad_id = pad_id
def __init__(
self,
manifest_filepath: Union[str, 'pathlib.Path'],
n_segments: int,
max_duration: Optional[float] = None,
min_duration: Optional[float] = None,
trim: Optional[bool] = False,
truncate_to: Optional[int] = 1,
):
"""
See above AudioDataset for details on dataset and manifest formats.
Unlike the regular AudioDataset, which samples random segments from each audio array as an example,
SplicedAudioDataset concatenates all audio arrays together and indexes segments as examples. This way,
the model sees more data (about 9x for LJSpeech) per epoch.
Note: this class is not recommended to be used in validation.
Args:
manifest_filepath (str, Path): Path to manifest json as described above. Can be comma-separated paths
such as "train_1.json,train_2.json" which is treated as two separate json files.
n_segments (int): The length of audio in samples to load. For example, given a sample rate of 16kHz, and
n_segments=16000, a random 1 second section of audio from the clip will be loaded. The section will
be randomly sampled everytime the audio is batched. Can be set to -1 to load the entire audio.
max_duration (float): If audio exceeds this length in seconds, it is filtered from the dataset.
Defaults to None, which does not filter any audio.
min_duration(float): If audio is less than this length in seconds, it is filtered from the dataset.
Defaults to None, which does not filter any audio.
trim (bool): Whether to use librosa.effects.trim on the audio clip
truncate_to (int): Ensures that the audio segment returned is a multiple of truncate_to.
Defaults to 1, which does no truncating.
"""
assert n_segments > 0
collection = collections.ASRAudioText(
manifests_files=manifest_filepath.split(','),
parser=parsers.make_parser(),
min_duration=min_duration,
max_duration=max_duration,
)
self.trim = trim
self.n_segments = n_segments
self.truncate_to = truncate_to
self.samples = []
for index in range(len(collection)):
example = collection[index]
with sf.SoundFile(example.audio_file, 'r') as f:
samples = f.read(dtype='float32').transpose()
self.samples.append(samples)
self.samples = np.concatenate(self.samples, axis=0)
self.samples = self.samples[:self.samples.shape[0] -
(self.samples.shape[0] % self.n_segments),
...]
def __init__(self, manifest_path=None, rng=None, audio_tar_filepaths=None, shuffle_n=128, shift_impulse=False):
self._manifest = collections.ASRAudioText(manifest_path, parser=parsers.make_parser([]), index_by_file_id=True)
self._audiodataset = None
self._tarred_audio = False
self._shift_impulse = shift_impulse
self._data_iterator = None
if audio_tar_filepaths:
self._tarred_audio = True
self._audiodataset = AugmentationDataset(manifest_path, audio_tar_filepaths, shuffle_n)
self._data_iterator = iter(self._audiodataset)
self._rng = random.Random() if rng is None else rng
def __init__(
self,
manifest_filepath: Union[str, "pathlib.Path"],
n_segments: int,
max_duration: Optional[float] = None,
min_duration: Optional[float] = None,
trim: Optional[bool] = False,
truncate_to: Optional[int] = 1,
):
"""
Mostly compliant with nemo.collections.asr.data.datalayers.AudioToTextDataset except it only returns Audio
without text. Dataset that loads tensors via a json file containing paths to audio files, transcripts, and
durations (in seconds). Each new line is a different sample. Note that text is required, but is ignored for
AudioDataset. Example below:
{"audio_filepath": "/path/to/audio.wav", "text_filepath":
"/path/to/audio.txt", "duration": 23.147}
...
{"audio_filepath": "/path/to/audio.wav", "text": "the
transcription", "offset": 301.75, "duration": 0.82, "utt":
"utterance_id", "ctm_utt": "en_4156", "side": "A"}
Args:
manifest_filepath (str, Path): Path to manifest json as described above. Can be comma-separated paths
such as "train_1.json,train_2.json" which is treated as two separate json files.
n_segments (int): The length of audio in samples to load. For example, given a sample rate of 16kHz, and
n_segments=16000, a random 1 second section of audio from the clip will be loaded. The section will
be randomly sampled everytime the audio is batched. Can be set to -1 to load the entire audio.
max_duration (float): If audio exceeds this length in seconds, it is filtered from the dataset.
Defaults to None, which does not filter any audio.
min_duration(float): If audio is less than this length in seconds, it is filtered from the dataset.
Defaults to None, which does not filter any audio.
trim (bool): Whether to use librosa.effects.trim on the audio clip
truncate_to (int): Ensures that the audio segment returned is a multiple of truncate_to.
Defaults to 1, which does no truncating.
"""
self.collection = collections.ASRAudioText(
manifests_files=manifest_filepath.split(","),
parser=parsers.make_parser(),
min_duration=min_duration,
max_duration=max_duration,
)
self.trim = trim
self.n_segments = n_segments
self.truncate_to = truncate_to
