audioDataAugmenter
Augment audio data
Description
Enlarge your audio dataset using audio-specific augmentation techniques like pitch shifting, time-scale modification, time shifting, noise addition, and volume control. You can create cascaded or parallel augmentation pipelines to apply multiple algorithms deterministically or probabilistically.
Creation
Description
aug = audioDataAugmenter()
aug = audioDataAugmenter(Name,Value)aug using one or more name-value
pair arguments.
Properties
Object Functions
addAugmentationMethod | Add custom augmentation method |
removeAugmentationMethod | Remove custom augmentation method |
augment | Augment audio data |
Examples
Apply Random Sequential Augmentations
Read in an audio signal and listen to it.
[audioIn,fs] = audioread("Counting-16-44p1-mono-15secs.wav");
sound(audioIn,fs)Create an audioDataAugmenter object that applies time stretching, volume control, and time shifting in cascade. Apply each of the augmentations with 80% probability. Set NumAugmentations to 5 to output five independently augmented signals. To skip pitch shifting and noise addition for each augmentation, set the respective probabilities to 0. Define parameter ranges for each relevant augmentation algorithm.
augmenter = audioDataAugmenter( ... "AugmentationMode","sequential", ... "NumAugmentations",5, ... ... "TimeStretchProbability",0.8, ... "SpeedupFactorRange", [1.3,1.4], ... ... "PitchShiftProbability",0, ... ... "VolumeControlProbability",0.8, ... "VolumeGainRange",[-5,5], ... ... "AddNoiseProbability",0, ... ... "TimeShiftProbability",0.8, ... "TimeShiftRange", [-500e-3,500e-3])
augmenter =
audioDataAugmenter with properties:
AugmentationMode: "sequential"
AugmentationParameterSource: 'random'
NumAugmentations: 5
TimeStretchProbability: 0.8000
SpeedupFactorRange: [1.3000 1.4000]
PitchShiftProbability: 0
VolumeControlProbability: 0.8000
VolumeGainRange: [-5 5]
AddNoiseProbability: 0
TimeShiftProbability: 0.8000
TimeShiftRange: [-0.5000 0.5000]
Call augment on the audio to create 5 augmentations. The augmented audio is returned in a table with variables Audio and AugmentationInfo. The number of rows in the table is defined by NumAugmentations.
data = augment(augmenter,audioIn,fs)
data=5×2 table
Audio AugmentationInfo
_________________ ________________
{685056x1 double} [1x1 struct]
{685056x1 double} [1x1 struct]
{505183x1 double} [1x1 struct]
{685056x1 double} [1x1 struct]
{490728x1 double} [1x1 struct]
In the current augmentation pipeline, augmentation parameters are assigned randomly from within the specified ranges. To determine the exact parameters used for an augmentation, inspect AugmentationInfo.
augmentationToInspect = 
4;
data.AugmentationInfo(augmentationToInspect)ans = struct with fields:
SpeedupFactor: 1
VolumeGain: 4.3399
TimeShift: 0.4502
Listen to the augmentation you are inspecting. Plot time representation of the original and augmented signals.
augmentation = data.Audio{augmentationToInspect};
sound(augmentation,fs)
t = (0:(numel(audioIn)-1))/fs;
taug = (0:(numel(augmentation)-1))/fs;
plot(t,audioIn,taug,augmentation)
legend("Original Audio","Augmented Audio")
ylabel("Amplitude")
xlabel("Time (s)")
Apply Specified Sequential Augmentations
Read in an audio signal and listen to it.
[audioIn,fs] = audioread("Counting-16-44p1-mono-15secs.wav");
sound(audioIn,fs)Create an audioDataAugmenter object that applies time stretching, pitch shifting, and noise corruption in cascade. Specify the time stretch speedup factors as 0.9, 1.1, and 1.2. Specify the pitch shifting in semitones as -2, -1, 1, and 2. Specify the noise corruption SNR as 10 dB and 15 dB.
augmenter = audioDataAugmenter( ... "AugmentationMode","sequential", ... "AugmentationParameterSource","specify", ... "SpeedupFactor",[0.9,1.1,1.2], ... "ApplyTimeStretch",true, ... "ApplyPitchShift",true, ... "SemitoneShift",[-2,-1,1,2], ... "SNR",[10,15], ... "ApplyVolumeControl",false, ... "ApplyTimeShift",false)
augmenter =
audioDataAugmenter with properties:
AugmentationMode: "sequential"
AugmentationParameterSource: "specify"
ApplyTimeStretch: 1
SpeedupFactor: [0.9000 1.1000 1.2000]
ApplyPitchShift: 1
SemitoneShift: [-2 -1 1 2]
ApplyVolumeControl: 0
ApplyAddNoise: 1
SNR: [10 15]
ApplyTimeShift: 0
Call augment on the audio to create 24 augmentations. The augmentations represent every combination of the specified augmentation parameters ().
data = augment(augmenter,audioIn,fs)
data=24×2 table
Audio AugmentationInfo
_________________ ________________
{761243x1 double} [1x1 struct]
{622888x1 double} [1x1 struct]
{571263x1 double} [1x1 struct]
{761243x1 double} [1x1 struct]
{622888x1 double} [1x1 struct]
{571263x1 double} [1x1 struct]
{761243x1 double} [1x1 struct]
{622888x1 double} [1x1 struct]
{571263x1 double} [1x1 struct]
{761243x1 double} [1x1 struct]
{622888x1 double} [1x1 struct]
{571263x1 double} [1x1 struct]
{761243x1 double} [1x1 struct]
{622888x1 double} [1x1 struct]
{571263x1 double} [1x1 struct]
{761243x1 double} [1x1 struct]
⋮
You can check the parameter configuration of each augmentation using the AugmentationInfo table variable.
augmentationToInspect = 
1;
data.AugmentationInfo(augmentationToInspect)ans = struct with fields:
SpeedupFactor: 0.9000
SemitoneShift: -2
SNR: 10
Listen to the augmentation you are inspecting. Plot the time-domain representation of the original and augmented signals.
augmentation = data.Audio{augmentationToInspect};
sound(augmentation,fs)
t = (0:(numel(audioIn)-1))/fs;
taug = (0:(numel(augmentation)-1))/fs;
plot(t,audioIn,taug,augmentation)
legend("Original Audio","Augmented Audio")
ylabel("Amplitude")
xlabel("Time (s)")
Apply Random Independent Augmentations
Read in an audio signal and listen to it.
[audioIn,fs] = audioread("Counting-16-44p1-mono-15secs.wav");Create an audioDataAugmenter object that applies noise corruption, and time shifting in parallel branches. For the noise corruption branch, randomly apply noise with an SNR in the range 0 dB to 20 dB. For the time shifting branch, randomly apply time shifting in the range -300 ms to 300 ms. Apply augmentation 2 times for each branch, for 4 total augmentations.
augmenter = audioDataAugmenter( ... "AugmentationMode","independent", ... "AugmentationParameterSource","random", ... "NumAugmentations",2, ... "ApplyTimeStretch",false, ... "ApplyPitchShift",false, ... "ApplyVolumeControl",false, ... "SNRRange",[0,20], ... "TimeShiftRange",[-300e-3,300e-3])
augmenter =
audioDataAugmenter with properties:
AugmentationMode: "independent"
AugmentationParameterSource: "random"
NumAugmentations: 2
ApplyTimeStretch: 0
ApplyPitchShift: 0
ApplyVolumeControl: 0
ApplyAddNoise: 1
SNRRange: [0 20]
ApplyTimeShift: 1
TimeShiftRange: [-0.3000 0.3000]
Call augment on the audio to create 3 augmentations.
data = augment(augmenter,audioIn,fs);
You can check the parameter configuration of each augmentation using the AugmentatioInfo table variable.
augmentationToInspect = 
4;
data.AugmentationInfo{augmentationToInspect}ans = struct with fields:
TimeShift: 0.0016
Listen to the audio you are inspecting. Plot the time-domain representation of the original and augmented signals.
augmentation = data.Audio{augmentationToInspect};
sound(augmentation,fs)
t = (0:(numel(audioIn)-1))/fs;
taug = (0:(numel(augmentation)-1))/fs;
plot(t,audioIn,taug,augmentation)
legend("Original Audio","Augmented Audio")
ylabel("Amplitude")
xlabel("Time (s)")
Apply Specified Independent Augmentations
Read in an audio signal and listen to it.
[audioIn,fs] = audioread("Counting-16-44p1-mono-15secs.wav");Create an audioDataAugmenter object that applies volume control, noise corruption, and time shifting in parallel branches.
augmenter = audioDataAugmenter( ... "AugmentationMode","independent", ... "AugmentationParameterSource","specify", ... "ApplyTimeStretch",false, ... "ApplyPitchShift",false, ... "VolumeGain",2, ... "SNR",0, ... "TimeShift",2)
augmenter =
audioDataAugmenter with properties:
AugmentationMode: "independent"
AugmentationParameterSource: "specify"
ApplyTimeStretch: 0
ApplyPitchShift: 0
ApplyVolumeControl: 1
VolumeGain: 2
ApplyAddNoise: 1
SNR: 0
ApplyTimeShift: 1
TimeShift: 2
Call augment on the audio to create 3 augmentations.
data = augment(augmenter,audioIn,fs)
data=3×2 table
Audio AugmentationInfo
_________________ ________________
{685056x1 double} {1x1 struct}
{685056x1 double} {1x1 struct}
{685056x1 double} {1x1 struct}
You can check the parameter configuration of each augmentation using the AugmentatioInfo table variable.
augmentationToInspect = 
3;
data.AugmentationInfo{augmentationToInspect}ans = struct with fields:
TimeShift: 2
Listen to the audio you are inspecting. Plot the time-domain representations of the original and augmented signals.
augmentation = data.Audio{augmentationToInspect};
sound(augmentation,fs)
t = (0:(numel(audioIn)-1))/fs;
taug = (0:(numel(augmentation)-1))/fs;
plot(t,audioIn,taug,augmentation)
legend("Original Audio","Augmented Audio")
ylabel("Amplitude")
xlabel("Time (s)")
Augment Audio Dataset
The audioDataAugmenter supports multiple workflows for augmenting your datastore, including:
Offline augmentation
Augmentation using tall arrays
Augmentation using transform datastores
In each workflow, begin by creating an audio datastore to point to your audio data. In this example, you create an audio datastore that points to audio samples included with Audio Toolbox™. Count the number of files in the dataset.
folder = fullfile(matlabroot,"toolbox","audio","samples"); ADS = audioDatastore(folder)
ADS =
audioDatastore with properties:
Files: {
' ...\matlab\toolbox\audio\samples\Ambiance-16-44p1-mono-12secs.wav';
' ...\matlab\toolbox\audio\samples\AudioArray-16-16-4channels-20secs.wav';
' ...\toolbox\audio\samples\ChurchImpulseResponse-16-44p1-mono-5secs.wav'
... and 26 more
}
AlternateFileSystemRoots: {}
OutputDataType: 'double'
Labels: {}
numFilesInDataset = numel(ADS.Files)
numFilesInDataset = 29
Create an audioDataAugmenter that applies random sequential augmentations. Set NumAugmentations to 2.
aug = audioDataAugmenter('NumAugmentations',2)aug =
audioDataAugmenter with properties:
AugmentationMode: 'sequential'
AugmentationParameterSource: 'random'
NumAugmentations: 2
TimeStretchProbability: 0.5000
SpeedupFactorRange: [0.8000 1.2000]
PitchShiftProbability: 0.5000
SemitoneShiftRange: [-2 2]
VolumeControlProbability: 0.5000
VolumeGainRange: [-3 3]
AddNoiseProbability: 0.5000
SNRRange: [0 10]
TimeShiftProbability: 0.5000
TimeShiftRange: [-0.0050 0.0050]
Offline Augmentation
To augment the audio dataset, create two augmentations of each file and then write the augmentations as WAV files.
while hasdata(ADS) [audioIn,info] = read(ADS); data = augment(aug,audioIn,info.SampleRate); [~,fn] = fileparts(info.FileName); for i = 1:size(data,1) augmentedAudio = data.Audio{i}; % If augmentation caused an audio signal to have values outside of -1 and 1, % normalize the audio signal to avoid clipping when writing. if max(abs(augmentedAudio),[],'all')>1 augmentedAudio = augmentedAudio/max(abs(augmentedAudio),[],'all'); end audiowrite(sprintf('%s_aug%d.wav',fn,i),augmentedAudio,info.SampleRate) end end
Create an audioDatastore that points to the augmented dataset and confirm that the number of files in the dataset is double the original number of files.
augmentedADS = audioDatastore(pwd)
augmentedADS =
audioDatastore with properties:
Files: {
' ...\Examples\audio-ex28074079\Ambiance-16-44p1-mono-12secs_aug1.wav';
' ...\Examples\audio-ex28074079\Ambiance-16-44p1-mono-12secs_aug2.wav';
' ...\Examples\audio-ex28074079\AudioArray-16-16-4channels-20secs_aug1.wav'
... and 55 more
}
AlternateFileSystemRoots: {}
OutputDataType: 'double'
Labels: {}
numFilesInAugmentedDataset = numel(augmentedADS.Files)
numFilesInAugmentedDataset = 58
Augment Using Tall Arrays
When augmenting a dataset using tall arrays, the input data to the augmenter should be sampled at a consistent rate. Subset the original audio dataset to only include files with a sample rate of 44.1 kHz. Most datasets are already cleaned to have a consistent sample rate.
keepFile = cellfun(@(x)contains(x,'44p1'),ADS.Files);
ads44p1 = subset(ADS,keepFile);
fs = 44.1e3;Convert the audio datastore to a tall array. tall arrays are evaluated only when you request them explicitly using gather. MATLAB® automatically optimizes the queued calculations by minimizing the number of passes through the data. If you have the Parallel Computing Toolbox™, you can spread the calculations across multiple machines. The audio data is represented as an M-by-1 tall cell array, where M is the number of files in the audio datastore.
adsTall = tall(ads44p1)
Starting parallel pool (parpool) using the 'local' profile ...
Connected to the parallel pool (number of workers: 6).
adsTall =
M×1 tall cell array
{ 539648×1 double}
{ 227497×1 double}
{ 8000×1 double}
{ 685056×1 double}
{ 882688×2 double}
{1115760×2 double}
{ 505200×2 double}
{3195904×2 double}
: :
: :
Define a cellfun function so that augmentation is applied to each cell of the tall array. Call gather to evaluate the tall array.
augTall = cellfun(@(x)augment(aug,x,fs),adsTall,"UniformOutput",false);
augmentedDataset = gather(augTall)Evaluating tall expression using the Parallel Pool 'local': - Pass 1 of 1: Completed in 1 min 34 sec Evaluation completed in 1 min 34 sec
augmentedDataset=12×1 cell array
{2×2 table}
{2×2 table}
{2×2 table}
{2×2 table}
{2×2 table}
{2×2 table}
{2×2 table}
{2×2 table}
{2×2 table}
{2×2 table}
{2×2 table}
{2×2 table}
The augmented dataset is returned as a numFiles-by-1 cell array, where numFiles is the number of files in the datastore. Each element of the cell array is a numAugmentationsPerFile-by-2 table, where numAugmentationsPerFile is the number of augmentations returned per file.
numFiles = numel(augmentedDataset)
numFiles = 12
numAugmentationsPerFile = size(augmentedDataset{1},1)numAugmentationsPerFile = 2
Augment Using Transform Datastore
You can perform online data augmentation while you train your machine learning application using a transform datastore. Call transform to create a new datastore that applies data augmentation while reading.
transformADS = transform(ADS,@(x,info)augment(aug,x,info),'IncludeInfo',true)transformADS =
TransformedDatastore with properties:
UnderlyingDatastore: [1×1 audioDatastore]
Transforms: {@(x,info)augment(aug,x,info)}
IncludeInfo: 1
Call read to return the augmented first file from the transform datastore.
augmentedRead = read(transformADS)
augmentedRead=2×2 table
Audio AugmentationInfo
_________________ ________________
{539648×1 double} [1×1 struct]
{586683×1 double} [1×1 struct]
Add Custom Augmentation Method
You can expand the capabilities of audioDataAugmenter by adding custom augmentation methods.
Read in an audio signal and listen to it.
[audioIn,fs] = audioread('Counting-16-44p1-mono-15secs.wav');
sound(audioIn,fs)Create an audioDataAugmenter object. Set the probability of applying white noise to 0.
augmenter = audioDataAugmenter('AddNoiseProbability',0)augmenter =
audioDataAugmenter with properties:
AugmentationMode: 'sequential'
AugmentationParameterSource: 'random'
NumAugmentations: 1
TimeStretchProbability: 0.5000
SpeedupFactorRange: [0.8000 1.2000]
PitchShiftProbability: 0.5000
SemitoneShiftRange: [-2 2]
VolumeControlProbability: 0.5000
VolumeGainRange: [-3 3]
AddNoiseProbability: 0
TimeShiftProbability: 0.5000
TimeShiftRange: [-0.0050 0.0050]
Specify a custom augmentation algorithm that applies pink noise. The AddPinkNoise algorithm is added to the augmenter properties.
algorithmName = 'AddPinkNoise'; algorithmHandle = @(x)x+pinknoise(size(x),'like',x); addAugmentationMethod(augmenter,algorithmName,algorithmHandle) augmenter
augmenter =
audioDataAugmenter with properties:
AugmentationMode: 'sequential'
AugmentationParameterSource: 'random'
NumAugmentations: 1
TimeStretchProbability: 0.5000
SpeedupFactorRange: [0.8000 1.2000]
PitchShiftProbability: 0.5000
SemitoneShiftRange: [-2 2]
VolumeControlProbability: 0.5000
VolumeGainRange: [-3 3]
AddNoiseProbability: 0
TimeShiftProbability: 0.5000
TimeShiftRange: [-0.0050 0.0050]
AddPinkNoiseProbability: 0.5000
Set the probability of adding pink noise to 1.
augmenter.AddPinkNoiseProbability = 1
augmenter =
audioDataAugmenter with properties:
AugmentationMode: 'sequential'
AugmentationParameterSource: 'random'
NumAugmentations: 1
TimeStretchProbability: 0.5000
SpeedupFactorRange: [0.8000 1.2000]
PitchShiftProbability: 0.5000
SemitoneShiftRange: [-2 2]
VolumeControlProbability: 0.5000
VolumeGainRange: [-3 3]
AddNoiseProbability: 0
TimeShiftProbability: 0.5000
TimeShiftRange: [-0.0050 0.0050]
AddPinkNoiseProbability: 1
Augment the original signal and listen to the result. Inspect parameters of the augmentation algorithms applied.
data = augment(augmenter,audioIn,fs);
sound(data.Audio{1},fs)
data.AugmentationInfo(1)ans = struct with fields:
SpeedupFactor: 1
SemitoneShift: 0
VolumeGain: 2.4803
TimeShift: -0.0022
AddPinkNoise: 'Applied'
Plot the mel spectrograms of the original and augmented signals.
melSpectrogram(audioIn,fs)
title('Original Signal')
melSpectrogram(data.Audio{1},fs)
title('Augmented Signal')
Algorithms
The audioDataAugmenter object enables you to configure your augmentation
pipeline as deterministic or probabilistic using the AugmentationParameterSource property. You can also choose to apply the
augmentations in series or in parallel using the AugmentationMode
property. The following sections describe the pipelines you can create and the applicable
properties for each architecture.
Random Sequential Augmentations
To define your augmentation as a sequence of probabilistically applied augmentations,
set AugmentationParameterSource to 'random' and AugmentationMode
to 'sequential'.
The order that augmentations are applied is always the same. If you specify custom algorithms, they are applied at the end of the sequence, in the order you specified them.
In this pipeline configuration, these parameters apply:
| Augmentation Method | Parameters |
|---|---|
| Stretch Time | |
| Shift Pitch | |
| Control Volume | |
| Add Noise | |
| Shift Time |
If you specify NumAugmentations
as greater than 1, then the object applies NumAugmentations parallel
random sequential augmentations. The probability of applying an augmentation, and the value
of any parameters that are probabilistically determined, are independent.
Specified Sequential Augmentations
To define your augmentation as a sequence of deterministically applied augmentations,
set AugmentationParameterSource to 'specify' and AugmentationMode
to 'sequential'.
The order that augmentations are applied is always the same. If you specify custom algorithms, they are applied at the end of the sequence, in the order you specified them.
In this pipeline configuration, these parameters apply:
| Augmentation Method | Parameters |
|---|---|
| Stretch Time | |
| Shift Pitch | |
| Control Volume | |
| Add Noise | |
| Shift Time |
If you specify an augmentation method as a vector, then each element of the vector creates a separate branch in the augmentation pipeline. For example, the following object creates an augmentation pipeline that results in four separate augmentations:
aug = audioDataAugmenter("AugmentationMode","sequential", ... "AugmentationParameterSource","specify", ... "SpeedupFactor",[0.8,1.2], ... "VolumeGain",[-3,-1])
aug =
audioDataAugmenter with properties:
AugmentationMode: "sequential"
AugmentationParameterSource: "specify"
ApplyTimeStretch: 1
SpeedupFactor: [0.8000 1.2000]
ApplyPitchShift: 1
SemitoneShift: -3
ApplyVolumeControl: 1
VolumeGain: [-3 -1]
ApplyAddNoise: 1
SNR: 5
ApplyTimeShift: 1
TimeShift: 0.0050
Random Independent Augmentations
To define your augmentation as independently applied augmentations with randomly
determined parameters, set AugmentationParameterSource to 'random' and AugmentationMode
to 'independent'.
In this pipeline configuration, these parameters apply:
| Augmentation Method | Parameters |
|---|---|
| Stretch Time | |
| Shift Pitch | |
| Control Volume | |
| Add Noise | |
| Shift Time |
If you specify NumAugmentations
as greater than 1, then the object applies NumAugmentations parallel
random independent augmentations. The value of any parameters that are probabilistically
determined are independent.
Specified Independent Augmentations
To define your augmentation as deterministically applied independent augmentations with
deterministic parameters, set AugmentationParameterSource to 'specify' and AugmentationMode
to 'independent'.
In this pipeline configuration, these parameters apply:
| Augmentation Method | Parameters |
|---|---|
| Stretch Time | |
| Shift Pitch | |
| Control Volume | |
| Add Noise | |
| Shift Time |
If you specify an augmentation method as a vector, then each element of the vector creates a separate branch in the augmentation pipeline. For example, the following object creates an augmentation pipeline that results in seven separate augmentations:
aug = audioDataAugmenter("AugmentationMode","independent", ... "AugmentationParameterSource","specify", ... "SpeedupFactor",[0.8,1.2], ... "VolumeGain",[-3,-1])
aug =
audioDataAugmenter with properties:
AugmentationMode: "independent"
AugmentationParameterSource: "specify"
ApplyTimeStretch: 1
SpeedupFactor: [0.8000 1.2000]
ApplyPitchShift: 1
SemitoneShift: -3
ApplyVolumeControl: 1
VolumeGain: [-3 -1]
ApplyAddNoise: 1
SNR: 5
ApplyTimeShift: 1
TimeShift: 0.0050
References
[1] Salamon, Justin, and Juan Pablo Bello. "Deep Convolutional Neural Networks and Data Augmentation for Environmental Sound Classification." IEEE Signal Processing Letters. Vol. 24, Issue 3, 2017.
See Also
audioFeatureExtractor | audioTimeScaler | shiftPitch | stretchAudio