Scaper¶
Scaper is a library for soundscape synthesis and augmentation.
For a quick introduction to using scaper, please refer to the Scaper tutorial. For a detailed description of scaper and its applications check out the scaper-paper:
Scaper: A library for soundscape synthesis and augmentation
J. Salamon, D. MacConnell, M. Cartwright, P. Li, and J. P. Bello
In IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), New Paltz, NY, USA, Oct. 2017.
Getting started¶
Installation instructions¶
Non-python dependencies¶
Scaper has one non-python dependency: - FFmpeg: https://ffmpeg.org/
If you are installing Scaper on Windows, you will also need: - SoX: http://sox.sourceforge.net/
On Linux/macOS SoX is replaced by [SoxBindings](https://github.com/pseeth/soxbindings) which is significantly faster, giving better runtime performance in Scaper. On these platforms SoxBindings is installed automatically when calling pip install scaper (see below).
On macOS ffmpeg can be installed using homebrew:
>>> brew install ffmpeg
On linux you can use your distribution’s package manager, e.g. on Ubuntu (15.04 “Vivid Vervet” or newer):
>>> sudo apt-get install ffmpeg
NOTE: on earlier versions of Ubuntu ffmpeg may point to a Libav binary which is not the correct binary. If you are using anaconda, you can install the correct version by calling:
>>> conda install -c conda-forge ffmpeg
Otherwise, you can obtain a static binary from the ffmpeg website.
On Windows you can use the provided installation binaries:
Installing Scaper¶
The simplest way to install scaper
is by using pip
, which will also install the required dependencies if needed.
To install scaper
using pip
, simply run
>>> pip install scaper
To install the latest version of scaper from source:
- Clone or pull the lastest version:
>>> git clone git@github.com:justinsalamon/scaper.git
- Install using pip to handle python dependencies:
>>> cd scaper
>>> pip install -e .
Scaper tutorial¶
Introduction¶
Welcome to the scaper tutorial! In this tutorial, we’ll explain how scaper works and show how to use scaper to synthesize soundscapes and generate corresponding annotation files.
Organize your audio files (source material)¶
Scaper creates new soundscapes by combining and transforming a set of existing audio files, which we’ll refer to as the source material. By combining/transforming the source material in different ways, scaper can create an infinite number of different soundscapes from the same source material. The source material is comprised of two groups of files: background files and foreground files:
- Background files: are used to create the background of the soundscape, and should contain audio material that is perceived as a single holistic sound which is more distant, ambiguous, and texture-like (e.g. the “hum” or “drone” of an urban environment, or “wind and rain” sounds in a natural environment). Importantly, background files should not contain salient sound events.
- Foreground files: are used to create sound events. Each foreground audio file should contain a single sound event (short or long) such as a car honk, an animal vocalization, continuous speech, a siren or an idling engine. Foreground files should be as clean as possible with no background noise and no silence before/after the sound event.
The source material must be organized as follows: at the top level, you need a
background
folder and a foreground
folder. Within each of these, scaper
expects a folder for each category (label) of sounds. Example categories of
background sounds include “street”, “forest” or “park”. Example categories of
foreground sounds include “speech”, “car_honk”, “siren”, “dog_bark”,
“bird_song”, “idling_engine”, etc. Within each category folder, scaper expects
WAV files of that category: background files should contain a single ambient
recording, and foreground files should contain a single sound event. The
filename of each audio file is not important as long as it ends with .wav
.
Here’s an example of a valid folder structure for the source material:
- foreground
- siren
- siren1.wav
- siren2.wav
- some_siren_sound.wav
- car_honk
- honk.wav
- beep.wav
- human_voice
- hello_world.wav
- shouting.wav
- background
- park
- central_park.wav
- street
- quiet_street.wav
- loud_street.wav
EXAMPLE SOURCE MATERIAL can be obtained by downloading the
scaper repository
(approx. 50mb). The audio can be found under scaper-master/tests/data/audio
(audio
contains two subfolders, background
and foreground
).
For the remainder of this tutorial, we’ll assume you’ve downloaded this material
and copied the audio
folder to your home directory. If you copy it somewhere
else (or use different source material), be sure to change the paths to the
foreground_folder
and background_folder
in the example code below.
Create a Scaper object¶
The first step is to create a Scaper
object:
import scaper
import os
path_to_audio = os.path.expanduser('~/audio')
soundscape_duration = 10.0
seed = 123
foreground_folder = os.path.join(path_to_audio, 'foreground')
background_folder = os.path.join(path_to_audio, 'background')
sc = scaper.Scaper(soundscape_duration, foreground_folder, background_folder)
sc.ref_db = -20
We need to supply three arguments to create a Scaper
object:
- The desired duration: all soundscapes generated by this Scaper object will have this duration.
- The path to the foreground folder.
- The path to the background folder.
If you’re not sure what the foreground and background folders are, please see Organize your audio files (source material).
Finally, we set the reference level sc.ref_db
, i.e. the loudnes of the
background, measured in LUFS. Later
when we add foreground events, we’ll have to specify an snr
(signal-to-noise ratio) value, i.e. by how many decibels (dB) should the foreground event
be louder (or softer) with respect to the background level specified by
sc.ref_db
.
Seeding the Scaper object for reproducibility¶
A further argument can be specified to the Scaper
object:
- The random state: this can be either a numpy.random.RandomState object or an integer. In the latter case, a random state will be constructed. The random state is what will be used for drawing from any distributions. If the audio kept in all of the folders is exactly the same and the random state is fixed between runs, the same soundscape will be generated both times. If you don’t define any random state or set seed to None, runs will be random and not reproducible. You can use np.random.get_state() to reproduce the run after the fact by recording the seed that was used somewhere.
This can be specified like so (e.g. for a random seed of 123):
seed = 123
sc = scaper.Scaper(soundscape_duration, foreground_folder, background_folder,
random_state=seed)
sc.ref_db = -20
If the random state is not specified, it defaults to the old behavior which just uses
the RandomState used by np.random. You can also set the random state after creation
via Scaper.set_random_state
. Alternatively, you can set the random state directly:
import numpy as np
seed = np.random.RandomState(123)
sc = scaper.Scaper(soundscape_duration, foreground_folder, background_folder,
random_state=seed)
sc.ref_db = -20
Adding a background and foreground sound events¶
Adding a background¶
Next, we can optionally add a background track to our soundscape:
sc.add_background(label=('const', 'park'),
source_file=('choose', []),
source_time=('const', 0))
To add a background we have to specify:
label
: the label (category) of background, which has to match the name of one of the subfolders in our background folder (in our example “park” or “street”).source_file
: the path to the specific audio file to be used.source_time
: the time in the source file from which to start the background.
Note how in the example above we do not specify these values directly by providing strings or floats, but rather we provide each arugment with a tuple. These tuples are called distribution tuples and are used in scaper for specifying all sound event parameters. Let’s explain:
Distribution tuples¶
One of the powerful things about scaper is that it allows you to define a soundscape
in a probabilistic way. That is, rather than specifying constant (hard coded) values for each
sound event, you can specify a distribution of values to sample from. Later on,
when we call sc.generate()
, a soundscape will be “instantiated” by sampling a value
for each distribution tuple in each sound event (foreground and background). Every time
we call sc.generate()
, a new value will be sampled for each distribution tuple,
resulting in a different soundscape.
The distribution tuples currently supported by scaper are:
('const', value)
: a constant, given byvalue
.('choose', list)
: uniformly sample from a finite set of values given bylist
.('uniform', min, max)
: sample from a uniform distribution betweenmin
andmax
.('normal', mean, std)
: sample from a normal distribution with meanmean
and standard deviationstd
.('truncnorm', mean, std, min, max)
: sample from a truncated normal distribution with meanmean
and standard deviationstd
, limited to values betweenmin
andmax
.
Special cases: the label
and source_file
parameters in sc.add_background()
(and as we’ll see later sc.add_event()
as well) must be specified using
either the const
or choose
distribution tuples. When using choose
, these
two parameters (and only these) can also accept a special version of the choose
tuple
in the form ('choose', [])
, i.e. with an empty list. In this case, scaper will
use the file structure in the foreground and background folders to automatically populate
the list with all valid labels (in the case of the label
parameter) and all valid
filenames (in the case of the source_file
parameter).
Adding a foreground sound event¶
Next, we can add foreground sound events. Let’s add one to start with:
sc.add_event(label=('const', 'siren'),
source_file=('choose', []),
source_time=('const', 0),
event_time=('uniform', 0, 9),
event_duration=('truncnorm', 3, 1, 0.5, 5),
snr=('normal', 10, 3),
pitch_shift=('uniform', -2, 2),
time_stretch=('uniform', 0.8, 1.2))
A foreground sound event requires several additional parameters compared to a background event. The full set of parameters is:
label
: the label (category) of foreground event, which has to match the name of one of the subfolders in our foreground folder (in our example “siren”, “car_honk” or “human_voice”).source_file
: the path to the specific audio file to be used.source_time
: the time in the source file from which to start the event.event_time
: the start time of the event in the synthesized soundscape.event_duration
: the duration of the event in the synthesized soundscape.snr
: the signal-to-noise ratio (in LUFS) compared to the background. In other words, how many dB above or below the background should this sound event be percieved.
Scaper also supports on-the-fly augmentation of sound events, that is, applying audio transformations to the sound events in order to increase the variability of the resulting soundscape. Currently, the supported transformations include pitch shifting and time stretching:
pitch_shift
: the number of semitones (can be fractional) by which to shift the sound up or down.time_stretch
: the factor by which to stretch the sound event. Factors <1 will make the event shorter, and factors >1 will make it longer.
If you do not wish to apply any transformations, these latter two parameters
(and only these) also accept None
instead of a distribution tuple.
So, going back to the example code above, we’re adding a siren sound event,
the specific audio file to use will be chosen randomly from all available siren
audio files in the foreground/siren
subfolder, the event will start at time
0 of the source file, and be “pasted” into the synthesized soundscape anywhere
between times 0 and 9 chosen uniformly. The event duration will be randomly
chosen from a truncated normal distribution with a mean of 3 seconds, standard
deviation of 1 second, and min/max values of 0.5 and 5 seconds respectively.
The loudness with respect to the background will be chosen from a normal
distribution with mean 10 dB and standard deviation of 3 dB. Finally, the pitch
of the sound event will be shifted by a value between -2 and 2 semitones
chosen uniformly within that range, and will be stretched (or condensed) by a
factor chosen uniformly between 0.8 and 1.2.
Let’s add a couple more events:
for _ in range(2):
sc.add_event(label=('choose', []),
source_file=('choose', []),
source_time=('const', 0),
event_time=('uniform', 0, 9),
event_duration=('truncnorm', 3, 1, 0.5, 5),
snr=('normal', 10, 3),
pitch_shift=None,
time_stretch=None)
Here we use a for loop to quickly add two sound events. The specific label and
source file for each event will be determined when we call sc.generate()
(coming up), and will change with each call to this function.
Synthesizing soundscapes¶
Up to this point, we have created a Scaper
object and added a background and
three foreground sound events, whose parameters are specified using distribution
tuples. Internally, this creates an event specification, i.e. a
probabilistically-defined list of sound events. To synthesize a soundscape,
we call the generate()
function:
audiofile = 'soundscape.wav'
jamsfile = 'soundscape.jams'
txtfile = 'soundscape.txt'
sc.generate(audiofile, jamsfile,
allow_repeated_label=True,
allow_repeated_source=True,
reverb=0.1,
disable_sox_warnings=True,
no_audio=False,
txt_path=txtfile)
This will instantiate the event specification by sampling specific parameter
values for every sound event from the distribution tuples stored in the
specification. Once all parameter values have been sampled, they are used by
scaper’s audio processing engine to compose the soundscape and save the
resulting audio to audiofile
.
But that’s not where it ends! Scaper will also generate an annotation file in
JAMS format which serves as the reference
annotation (also referred to as “ground truth”) for the generated soundscape.
Due to the flexibility of the JAMS
format scaper will store in the JAMS file, in addition to the actual sound
events, the probabilistic event specification (one for background events and one
for foreground events). The value
field of each observation in the JAMS file
will contain a dictionary with all instantiated parameter values. This allows
us to fully reconstruct the audio of a scaper soundscape from its JAMS annotation
using the scaper.generate_from_jams()
function (not discussed in this tutorial).
We can optionally provide generate()
a path to a text file
with the txt_path
parameter. If provided, scaper will also save a simplified
annotation of the soundscape in a tab-separated text file with three columns
for the start time, end time, and label of every foreground sound event (note that
the background is not stored in the simplified annotation!). The default
separator is a tab, for compatibility with the Audacity
label file format. The separator can be changed via generate()
’s txt_sep
parameter.
Synthesizing isolated events alongside the soundscape¶
We can also output the isolated foreground events and backgrounds alongside the soundscape.
This is especially useful for generating datasets that can be used to train and evaluate
source separation algorithms or models. To enable this, two additional arguments can be
given to generate()
and generate_from_jams()
:
save_isolated_events
: whether or not to save the audio corresponding to the to the isolated foreground events and backgrounds within the synthesized soundscape. In our example, there are three components - the background and the two foreground events.isolated_events_path
: the path where the audio corresponding to the isolated foreground events and backgrounds will be saved. If None (default) and save_isolated_events = True, the events are saved to <parentdir>/<audiofilename>_events/, where <parentdir> is the parent folder of the soundscape audio file provided in the audiofile parameter in the example below:
audiofile = '~/scaper_output/mysoundscape.wav'
jamsfile = '~/scaper_output/mysoundscape.jams'
txtfile = '~/scaper_output/mysoundscape.txt'
sc.generate(audiofile, jamsfile,
allow_repeated_label=True,
allow_repeated_source=True,
reverb=None,
disable_sox_warnings=True,
no_audio=False,
txt_path=txtfile,
save_isolated_events=True)
The code above will produce the following directory structure:
~/scaper_output/mysoundscape.wav
~/scaper_output/mysoundscape.jams
~/scaper_output/mysoundscape.txt
~/scaper_output/mysoundscape_events/
background0_<label>.wav
foreground0_<label>.wav
foreground1_<label>.wav
The labels for each isolated event are determined after generate
is called.
If isolated_events_path
were specified, then it would produce:
~/scaper_output/mysoundscape.wav
~/scaper_output/mysoundscape.jams
~/scaper_output/mysoundscape.txt
<isolated_events_path>/
background0_<label>.wav
foreground0_<label>.wav
foreground1_<label>.wav
The audio of the isolated events is guaranteed to sum up to the soundscape audio if and
only if reverb
is None
! The audio of the isolated events as well as the audio
of the soundscape can be accessed directly via the jams file as follows:
import soundfile as sf
jam = jams.load(jams_file)
ann = jam.annotations.search(namespace='scaper')[0]
soundscape_audio, sr = sf.read(ann.sandbox.scaper.soundscape_audio_path)
isolated_event_audio_paths = ann.sandbox.scaper.isolated_events_audio_path
isolated_audio = []
for event_spec, event_audio in zip(ann, isolated_event_audio_paths):
# event_spec contains the event description, label, etc
# event_audio contains the path to the actual audio
# make sure the path matches the event description
assert event_spec.value['role'] in event_audio_path
assert event_spec.value['label'] in event_audio_path
isolated_audio.append(sf.read(event_audio)[0])
# the sum of the isolated audio should sum to the soundscape
assert sum(isolated_audio) == soundscape_audio
That’s it! For a more detailed example of automatically synthesizing 1000
soundscapes using a single Scaper
object, please see the Example: synthesizing 1000 soundscapes in one go.
Examples¶
Example: synthesizing 1000 soundscapes in one go¶
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 | import scaper
import numpy as np
# OUTPUT FOLDER
outfolder = 'audio/soundscapes/'
# SCAPER SETTINGS
fg_folder = 'audio/soundbank/foreground/'
bg_folder = 'audio/soundbank/background/'
n_soundscapes = 1000
ref_db = -50
duration = 10.0
min_events = 1
max_events = 9
event_time_dist = 'truncnorm'
event_time_mean = 5.0
event_time_std = 2.0
event_time_min = 0.0
event_time_max = 10.0
source_time_dist = 'const'
source_time = 0.0
event_duration_dist = 'uniform'
event_duration_min = 0.5
event_duration_max = 4.0
snr_dist = 'uniform'
snr_min = 6
snr_max = 30
pitch_dist = 'uniform'
pitch_min = -3.0
pitch_max = 3.0
time_stretch_dist = 'uniform'
time_stretch_min = 0.8
time_stretch_max = 1.2
# generate a random seed for this Scaper object
seed = 123
# create a scaper that will be used below
sc = scaper.Scaper(duration, fg_folder, bg_folder, random_state=seed)
sc.protected_labels = []
sc.ref_db = ref_db
# Generate 1000 soundscapes using a truncated normal distribution of start times
for n in range(n_soundscapes):
print('Generating soundscape: {:d}/{:d}'.format(n+1, n_soundscapes))
# reset the event specifications for foreground and background at the
# beginning of each loop to clear all previously added events
sc.reset_bg_spec()
sc.reset_fg_spec()
# add background
sc.add_background(label=('const', 'noise'),
source_file=('choose', []),
source_time=('const', 0))
# add random number of foreground events
n_events = np.random.randint(min_events, max_events+1)
for _ in range(n_events):
sc.add_event(label=('choose', []),
source_file=('choose', []),
source_time=(source_time_dist, source_time),
event_time=(event_time_dist, event_time_mean, event_time_std, event_time_min, event_time_max),
event_duration=(event_duration_dist, event_duration_min, event_duration_max),
snr=(snr_dist, snr_min, snr_max),
pitch_shift=(pitch_dist, pitch_min, pitch_max),
time_stretch=(time_stretch_dist, time_stretch_min, time_stretch_max))
# generate
audiofile = os.path.join(outfolder, "soundscape_unimodal{:d}.wav".format(n))
jamsfile = os.path.join(outfolder, "soundscape_unimodal{:d}.jams".format(n))
txtfile = os.path.join(outfolder, "soundscape_unimodal{:d}.txt".format(n))
sc.generate(audiofile, jamsfile,
allow_repeated_label=True,
allow_repeated_source=False,
reverb=0.1,
disable_sox_warnings=True,
no_audio=False,
txt_path=txtfile)
|
Contribute¶
Changes¶
Changelog¶
v1.6.5.rc0¶
- Added a new distirbution tuple:
("choose_weighted", list_of_options, probabilities)
, which supports weighted sampling:list_of_options[i]
is chosen with probabilityprobabilities[i]
.
v1.6.4¶
- Scaper.generate now accepts a new argument for controlling trade-off between speed and quality in pitch shifting and time stretching:
- quick_pitch_time: if True, both time stretching and pitch shifting will be applied in quick mode, which is much faster but has lower quality.
v1.6.3¶
- Scaper.generate now accepts two new optional arguments for controlling audio clipping and normalization:
- fix_clipping: if True and the soundscape audio is clipping, it will be peak normalized and all isolated events will be scaled accordingly.
- peak_normalization: if True, sounscape audio will be peak normalized regardless of whether it’s clipping or not and all isolated events will be scaled accordingly.
- All generate arguments are now documented in the scaper sandbox inside the JAMS annotation.
- Furthermore, we also document in the JAMS: the scale factor used for peak normalization, the change in ref_db, and the actual ref_db of the generated audio.
v1.6.2¶
- Switching from FFMpeg LUFS calculation to pyloudnorm for better performance: runtime is reduced by approximately 30%
- The loudness calculation between FFMpeg LUFS and pyloudnorm is slightly different so this version will generate marginally different audio data compared to previous versions: the change is not perceptible, but np.allclose() tests on audio from previous versions of Scaper may fail.
- This change updates the regression data for Scaper’s regression tests.
- This release used soxbindings 1.2.2 and pyloudnorm 0.1.0.
v1.6.1¶
- Trimming now happens on read, rather than after read. This prevents the entire file from being loaded into memory. This is helpful for long source audio files.
- Since the audio processing pipeline has changed, this version will generate marginally different audio data compared to previous versions: the change is not perceptible, but np.allclose() tests on audio from previous versions of Scaper may fail.
- This change updates the regression data for Scaper’s regression tests
v1.6.0¶
- Uses soxbindings when installing on Linux or MacOS, which results in better performance.
- Adds explicit support for Python 3.7 and 3.8. Drops support for Python 2.7 and 3.4.
v1.5.1¶
- Fixes a bug with fade in and out lengths are set to 0.
- This is the last version to support Python 2.7 and 3.4.
v1.5.0¶
- Scaper now returns the generated audio and annotations directly in memory, allowing you to skip any/all file I/O!
- Saving the audio and annotations to disk is still supported, but is now optional.
- While this update modifies the API of several functions, it should still be backwards compatible.
v1.4.0¶
- Operations on all files happen in-memory now, via new PySox features (build_array) and numpy operations for applying fades.
- Scaper is faster now due to the in-memory changes.
v1.3.9¶
- Fixed a bug where trim before generating soundscapes from a JAMS file with saving of isolated events resulted in incorrect soundscape audio.
v1.3.8¶
- Fixed a bug where _sample_trunc_norm returned an array in Scipy 1.5.1, but returns a scalar in Scipy 1.4.0.
v1.3.7¶
- Fixed a bug where time stretched events could have a negative start time if they were longer than the soundscape duration.
v1.3.6¶
- Use sox flag -s for time stretching (speech mode), gives better sounding results.
v1.3.5¶
- Fixed a bug where short backgrounds did not concatenate to fill the entire soundscape.
v1.3.4¶
- Fixed a bug where the soundscapes were off by one sample when generated. Fixes bug where generating from jams using a trimmed jams file was using the trimmed soundscape duration instead of the original duration.
- Added a field to the sandbox that keeps track of the original duration of the soundscape before any trimming is applied.
v1.3.3¶
- Fixed a bug with the format and subtype of audio files not being maintained in match_sample_length.
v1.3.2¶
- Fixed a bug with generating the file names when saving the isolated events. The idx for background and foreground events now increment separately.
v1.3.1¶
- Fixed a bug with generating docs on ReadTheDocs.
v1.3.0¶
- Source separation support! Add option to save isolated foreground events and background audio files.
- Makes pysoundfile a formal dependency.
- Seeding tests more robust.
v1.2.0¶
- Added a random_state parameter to Scaper object, which allows all runs to be perfectly reproducible given the same audio and the same random seed.
- Switched from numpydoc to napoleon for generating the documentation. Also switched Sphinx to the most recent version.
- Added functions to Scaper object that allow one to reset the foreground and background event specifications independently. This allows users to reuse the same Scaper object and generate multiple soundscapes.
- Added a function to Scaper that allows the user to set the random state after creation.
v1.1.0¶
- Added functionality which modifies a source_time distribution tuple according to the duration of the source and the duration of the event.
- This release alters behavior of Scaper compared to earlier versions.
v1.0.3¶
- Fix bug where temp files might not be closed if an error is raised
v1.0.2¶
- Store sample rate in output JAMS inside the scaper sandbox
v1.0.1¶
- Fix bug where estimated duration of time stretched event is different to actual duration leading to incorrect silence padding and sometimes incorrect soundscape duration (in audio samples).
v1.0.0¶
- Major revision
- Support jams>=0.3
- Switch from the sound_event to the scaper namespace.
- While the API remains compatible with previous versions, the change of underlying namespace breaks compatibility with jams files created using scaper for versions <1.0.0.
v0.2.1¶
- Fix bug related to creating temp files on Windows.
v0.2.0¶
#28: Improve LUFS calculation:
- Compute LUFS after initial processing (e.g. trimming, augmentation) of foreground and background events
- Self-concatenate short events (< 500 ms) to avoid ffmpeg constant of -70.0 LUFS
v0.1.2¶
- Fix markdown display on PyPi
v0.1.1¶
- Increases minimum version of pysox to 1.3.3 to prevent crashing on Windows
v0.1.0¶
- First release.