To install Soundata simply do:
pip install soundata
Soundata is easily imported into your Python code by:
import soundataPrint a list of all available dataset loaders by calling:
import soundata
print(soundata.list_datasets())To use a loader, (for example, urbansound8k) you need to initialize it by calling:
import soundata
dataset = soundata.initialize('urbansound8k', data_home='choose_where_data_live')You can indicate where the data would be stored and access by passing a path to data_home, as explained below. Now us8k is a Dataset object containing common methods, described in the following.
All dataset loaders in soundata have a download() function that allows the user to download the canonical <faq> version of the dataset (when available). When initializing a dataset it is important to correctly set up the directory where the dataset is going to be stored and retrieved.
- Downloading a dataset into the default folder:
In this first example,
data_homeis not specified. Thus, UrbanSound8K will be downloaded and retrieved from the default folder,sound_datasets, created in the user's root folder:import soundata dataset = soundata.initialize('urbansound8k') dataset.download() # Dataset is downloaded into "sound_datasets" folder inside user's root folder
- Downloading a dataset into a specified folder:
In the next example
data_homeis specified, so UrbanSound8K will be downloaded and retrieved from the specified location:dataset = soundata.initialize('urbansound8k', data_home='Users/johnsmith/Desktop') dataset.download() # Dataset is downloaded to John Smith's desktop
- Partially downloading a dataset
The
download()function allows to partially download a dataset. In other words, if applicable, the user can select which elements of the dataset they want to download. Each dataset has aREMOTESdictionary were all the available downloadable elements are listed.tau2019uashas different elements as seen in theREMOTESdictionary. You can specify a subset of these elements to download by passing thedownload()function a list of theREMOTESkeys that we are interested in via thepartial_downloadvariable.Example REMOTES
REMOTES = { "development.audio.1": download_utils.RemoteFileMetadata( filename="TAU-urban-acoustic-scenes-2019-development.audio.1.zip", url="https://zenodo.org/record/2589280/files/TAU-urban-acoustic-scenes-2019-development.audio.1.zip?download=1", checksum="aca4ebfd9ed03d5f747d6ba8c24bc728", ), "development.audio.2": download_utils.RemoteFileMetadata( filename="TAU-urban-acoustic-scenes-2019-development.audio.2.zip", url="https://zenodo.org/record/2589280/files/TAU-urban-acoustic-scenes-2019-development.audio.2.zip?download=1", checksum="c4f170408ce77c8c70c532bf268d7be0", ), "development.audio.3": download_utils.RemoteFileMetadata( filename="TAU-urban-acoustic-scenes-2019-development.audio.3.zip", url="https://zenodo.org/record/2589280/files/TAU-urban-acoustic-scenes-2019-development.audio.3.zip?download=1", checksum="c7214a07211f10f3250290d05e72c37e", ), ....
A partial download example for
tau2019uasdataset could be:dataset = soundata.initialize('tau2019uas') dataset.download(partial_download=['development.audio.1', 'development.audio.2']) # download only two remotes
- Downloading a multipart dataset
In some cases, datasets consist of multiple remote files that have to be extracted together locally to correctly recover the data. In those cases, remotes that need to be extracted together should be grouped in a list, so all the necessary files are downloaded at once (even in a partial download). An example of this is the fsd50k loader:
Example multipart REMOTES
REMOTES = { "FSD50K.dev_audio": [ download_utils.RemoteFileMetadata( filename="FSD50K.dev_audio.zip", url="https://zenodo.org/record/4060432/files/FSD50K.dev_audio.zip?download=1", checksum="c480d119b8f7a7e32fdb58f3ea4d6c5a", ), download_utils.RemoteFileMetadata( filename="FSD50K.dev_audio.z01", url="https://zenodo.org/record/4060432/files/FSD50K.dev_audio.z01?download=1", checksum="faa7cf4cc076fc34a44a479a5ed862a3", ), download_utils.RemoteFileMetadata( filename="FSD50K.dev_audio.z02", url="https://zenodo.org/record/4060432/files/FSD50K.dev_audio.z02?download=1", checksum="8f9b66153e68571164fb1315d00bc7bc", ), download_utils.RemoteFileMetadata( filename="FSD50K.dev_audio.z03", url="https://zenodo.org/record/4060432/files/FSD50K.dev_audio.z03?download=1", checksum="1196ef47d267a993d30fa98af54b7159", ), download_utils.RemoteFileMetadata( filename="FSD50K.dev_audio.z04", url="https://zenodo.org/record/4060432/files/FSD50K.dev_audio.z04?download=1", checksum="d088ac4e11ba53daf9f7574c11cccac9", ), download_utils.RemoteFileMetadata( filename="FSD50K.dev_audio.z05", url="https://zenodo.org/record/4060432/files/FSD50K.dev_audio.z05?download=1", checksum="81356521aa159accd3c35de22da28c7f", ), ], ...
Using the validate() method you can ensure that the files in our local copy of a dataset are identical to the canonical <faq> version of the dataset. The function computes the md5 checksum of every downloaded file to ensure it was downloaded correctly and isn't corrupted.
For big datasets: In future soundata versions, a random validation will be included. This improvement will reduce validation time for very big datasets.
You can choose a random clip from a dataset with the choice_clip() method.
Example Index
dataset = soundata.initialize('urbansed')
random_clip = dataset.choice_clip()
print(random_clip)
>>> Clip(
audio_path="/Users/theuser/sound_datasets/urbansed/audio/test/soundscape_test_bimodal73.wav",
clip_id="soundscape_test_bimodal73",
jams_path="/Users/mf3734/sound_datasets/urbansed/annotations/test/soundscape_test_bimodal73.jams",
txt_path="/Users/mf3734/sound_datasets/urbansed/annotations/test/soundscape_test_bimodal73.txt",
audio: The clips audio
* np.ndarray - audio signal
* float - sample rate,
events: The audio events
* annotations.Events - audio event object,
split: The data splits (e.g. train)
* str - split,
)You can also access specific clips by id. The available clip ids can be acessed by doing dataset.clip_ids. In the next example we take the first clip id, and then we retrieve its tags annotation.
dataset = soundata.initialize('urbansound8k')
ids = dataset.clip_ids # the list of urbansound8k's clip ids
clips = dataset.load_clips() # Load all clips in the dataset
example_clip = clips[ids[0]] # Get the first clip
# Accessing the clip's tags annotation
example_tags = example_clip.tags
print(example_tags)
>>>> Tags(confidence, labels, labels_unit)
print(example_tags.labels)
>>>> ['children_playing']You can also load a single clip without loading all clips in the dataset:
ids = dataset.clip_ids # the list of urbansound8k's clip ids
example_clip = dataset.clip(ids[0]) # load this particular clip
example_tags = example_clip.tags # Get the tags for the first clipAnnotations can also be accessed through load_*() methods which may be useful, for instance, when your data aren't available locally. If you specify the annotation's path, you can use the module's loading functions directly. Let's see an example.
Accessing annotations remotely example
# Load list of clip ids of the dataset
ids = dataset.clip_ids
# Load a single clip, specifying the remote location
example_clip = dataset.clip(ids[0], data_home='remote/data/path')
audio_path = example_clip.audio_path
print(audio_path)
>>> remote/data/path/audio/fold1/135776-2-0-49.wav
print(os.path.exists(audio_path))
>>> False
# Write code here to download the remote path, e.g., to a temporary file.
def my_downloader(remote_path):
# the contents of this function will depend on where your data lives, and how permanently you
# want the files to remain on your local machine. We point you to libraries handling common use cases below.
# for data you would download via scp, you could use the [scp](https://pypi.org/project/scp/) library
# for data on google drive, use [pydrive](https://pythonhosted.org/PyDrive/)
# for data on google cloud storage use [google-cloud-storage](https://pypi.org/project/google-cloud-storage/)
return local_path_to_downloaded_data
# Get path to where your data live
temp_path = my_downloader(audio_path)
# Accessing the clip audio
example_audio = dataset.load_audio(temp_path)soundata defines annotation-specific data classes such as Tags or Events. These data classes are meant to standarize the format for all loaders, so you can use the same code with different datasets. The list and descriptions of available annotation classes can be found in annotations.
Note
These classes are standarized to the point that the data allow for it. In some cases where the dataset has its own idiosyncrasies, the classes may be extended e.g. adding a customize, uncommon attribute.
In general, most datasets are a collection of clips, and in most cases each clip has an audio file along with annotations.
With the load_clips() method, all clips are loaded as a dictionary with the clip id as keys and clip objects as values. The clip objects include their respective audio and annotations, which are lazy-loaded on access to keep things speedy and memory efficient.
dataset = soundata.initialize('urbansound8k')
for key, clip in dataset.load_clips().items():
print(key, clip.audio_path)
>>>> soundscape_train_bimodal0 /Users/mf3734/sound_datasets/urbansed/audio/train/soundscape_train_bimodal0.wav
.....Alternatively, you can loop over the clip_ids list to directly access each clip in the dataset.
dataset = soundata.initialize('urbansound8k')
for clip_id in dataset.clip_ids:
print(clip_id, dataset.clip(clip_id).audio_path)
>>>> soundscape_train_bimodal0 /Users/mf3734/sound_datasets/urbansed/audio/train/soundscape_train_bimodal0.wav
.....If you wanted to use urbansound8k to evaluate the performance of an urban sound classifier, (in our case, random_classifier), and then split the scores based on the metadata, you could do the following:
soundata usage example
import sed_eval
import soundata
import numpy as np
from dcase_util.containers import MetaDataContainer, ProbabilityContainer
def random_classifier(classes):
return [np.random.random(1)[0] for c in classes]
# Evaluate on the full dataset
dataset = soundata.initialize('urbansound8k')
scores = {}
data = dataset.load_clips()
classes = np.unique([c for _, clip_data in data.items() for c in clip_data.tags.labels])
fold = 2 # Choose a fold to evaluate
ref_tags, est_tags, est_tag_probs = [], [], []
for id, clip in data.items():
if clip.fold == 2:
ref_tags.append({'filename': id, 'tags': clip.tags.labels[0]}) # Urbansound8k has one label per clip
probs = random_classifier(classes)
for c, p in zip(classes, probs):
est_tag_probs.append({'filename': id, 'label': c, 'probability': p},)
if p > 0.5: # Detection threshold of 0.5
est_tags.append({'filename': id, 'tags': [c]})
tag_evaluator = sed_eval.audio_tag.AudioTaggingMetrics(tags=MetaDataContainer(ref_tags).unique_tags)
tag_evaluator.evaluate(
reference_tag_list=MetaDataContainer(ref_tags),
estimated_tag_list=MetaDataContainer(est_tags),
estimated_tag_probabilities=ProbabilityContainer(est_tag_probs))This is the result of the example above:
Example result
print(tag_evaluator)
>>> Audio tagging metrics
========================================
Tags : 10
Evaluated units : 888
Overall metrics (micro-average)
======================================
F-measure
F-measure (F1) : 9.57 %
Precision : 9.57 %
Recall : 9.57 %
Equal error rate
Equal error rate (EER) : 51.01 %
Class-wise average metrics (macro-average)
======================================
F-measure
F-measure (F1) : 6.47 %
Precision : 7.54 %
Recall : 9.33 %
Equal error rate
Equal error rate (EER) : 50.95 %
Class-wise metrics
======================================
Tag | Nref Nsys | F-score Pre Rec | EER
----------------- | --------- --------- | --------- --------- --------- | ---------
air_conditioner | 100 419 | 19.3% 11.9 50.0 | 49.0%
car_horn | 42 227 | 4.5% 2.6 14.3 | 54.8%
children_playing | 100 126 | 9.7% 8.7 11.0 | 54.0%
dog_bark | 100 58 | 13.9% 19.0 11.0 | 47.1%
drilling | 100 31 | 9.2% 19.4 6.0 | 52.4%
engine_idling | 100 16 | 1.7% 6.2 1.0 | 50.0%
gun_shot | 35 7 | 0.0% 0.0 0.0 | 48.1%
jackhammer | 120 1 | 0.0% 0.0 0.0 | 52.5%
siren | 91 3 | 0.0% 0.0 0.0 | 51.6%
street_music | 100 0 | nan% nan 0.0 | 50.0%The following is a simple example of a generator that can be used to create a tensorflow Dataset.
soundata with tf.data.Dataset example
import soundata
import numpy as np
import tensorflow as tf
def data_generator(dataset_name):
# using the default data_home
dataset = soundata.initialize(dataset_name)
ids = dataset.clip_ids()
for clip_id in ids:
clip = dataset.clip(clip_id)
audio_signal, sample_rate = clip.audio
yield {
"audio": audio_signal.astype(np.float32),
"sample_rate": sample_rate,
"label": clip.tags.labels[0],
"metadata": {"clip_id": clip.clip_id, "fold": clip.fold}
}
dataset = tf.data.Dataset.from_generator(
data_generator('urbansound8k'),
{
"audio": tf.float32,
"sample_rate": tf.float32,
"label": tf.string,
"metadata": {'clip_id': tf.string, 'fold': tf.string}
}
)In future soundata versions, generators for Tensorflow and PyTorch will be included out-of-the-box.