Audio augmentations library for PyTorch for audio in the time-domain

Last update: Jan 08, 2023

Related tags

Overview

Audio Augmentations

Audio augmentations library for PyTorch for audio in the time-domain, with support for stochastic data augmentations as used often in self-supervised / contrastive learning.

Usage

We can define several audio augmentations, which will be applied sequentially to a raw audio waveform:

from audio_augmentations import *

audio, sr = torchaudio.load("tests/classical.00002.wav")

num_samples = sr * 5
transforms = [
    RandomResizedCrop(n_samples=num_samples),
    RandomApply([PolarityInversion()], p=0.8),
    RandomApply([Noise(min_snr=0.3, max_snr=0.5)], p=0.3),
    RandomApply([Gain()], p=0.2),
    RandomApply([HighLowPass(sample_rate=sr)], p=0.8),
    RandomApply([Delay(sample_rate=sr)], p=0.5),
    RandomApply([PitchShift(
        n_samples=num_samples,
        sample_rate=sr
    )], p=0.4),
    RandomApply([Reverb(sample_rate=sr)], p=0.3)
]

We can return either one or many versions of the same audio example:

transform = Compose(transforms=transforms)
transformed_audio =  transform(audio)
>> transformed_audio.shape[0] = 1

> transformed_audio.shape[0] = 4 ">

audio = torchaudio.load("testing/classical.00002.wav")
transform = ComposeMany(transforms=transforms, num_augmented_samples=4)
transformed_audio = transform(audio)
>> transformed_audio.shape[0] = 4

Similar to the torchvision.datasets interface, an instance of the Compose or ComposeMany class can be supplied to a torchaudio dataloaders that accept transform=.

Optional

Install WavAugment for reverberation / pitch shifting:

pip install git+https://github.com/facebookresearch/WavAugment

Cite

You can cite this work with the following BibTeX:

@misc{spijkervet_torchaudio_augmentations,
  doi = {10.5281/ZENODO.4748582},
  url = {https://zenodo.org/record/4748582},
  author = {Spijkervet,  Janne},
  title = {Spijkervet/torchaudio-augmentations},
  publisher = {Zenodo},
  year = {2021},
  copyright = {MIT License}
}

Comments

Delay augmentation on cuda

Hi. Currently the delay augmentation doesn't work on gpu since part of the signal is on cpu. I think making thebeginning tensor same as the audio tensor device should fix it. Thanks. https://github.com/Spijkervet/torchaudio-augmentations/blob/d044f9d020e12032ab9280acf5f34a337e72d212/torchaudio_augmentations/augmentations/delay.py#L31

opened by sidml 2
Correctness unit test would be great

For some transforms, we can test if the values are actually correct by manually computing the expected value. For example, PolarityInversion could be test with some tiny tensors like [[0.1, 0.5, -1.0]]. Reverse as well. Probably only those two? Still, it'd be better than not having any.

opened by keunwoochoi 2
Default value of `max_snr` in `Noise`

1.0 of SNR with signal and white noise would be a really heavily corrupted signal. Could we set it to be a little more reasonable value?

Related; it'd be great if one can hear some examples of the augmented result.

opened by keunwoochoi 2
End-to-end PitchShift transform tests

This merge requests adds end-to-end pitch transformation detection with librosa's pYIN pitch detection, to test if the applied transformation yields the expected pitch transposition.

opened by Spijkervet 0
Unittests

This adds various unittests and fixes to multi-channel input for Reverb, Pitch, Reverse and HighLowPass filter augmentations. It also removes Essentia as a dependency, and instead uses julius for IRR filtering.

opened by Spijkervet 0
import error

when i import torchaudio_augmentation

I got the error

RuntimeError : torchaudio.sox_effects.sox_effects.effect_names requires module: torchaudio._torchaudio

how can I deal with it?

opened by EavnJeong 0
Snr db
Hi, Thanks for the interesting work. Allow me to suggest this change for two reasons:

Expressing SNR in dB cancels the doubt there might be between power SNR and RMS SNR.

When sampling an SNR, it feels to me like it makes more sense to uniformly sample from the log scale of the dB than on the linear range. This way you ensure that your low SNR have as much chances as your high SNR.

I hope it makes sens. I'd be glad to discuss further about that.
opened by wesbz 0
sanity check for duration

In transforms where the duration may change, if the input audio is shorter than n_samples, the error message is not intuitive. I forgot but in some case, the multiprocessing-based dataloader silently died. Maybe it's worth checking it somewhere?

opened by keunwoochoi 0
Shapes are still a bit confusing

From ComposeMany.__call__(), is x also a ch, time shape 2-dim tensor? And I'm sure what would be the expected behavior by this function, especially the shape of the output.

opened by keunwoochoi 3

Releases(v0.2.3)

v0.2.3(Nov 15, 2021)
This version:

Removes WavAugment's pitch detection, and instead relies on the parallelisable torch-pitch-shift package.

Adds support for batched inputs (batch, channel, time).

It also adds more tests for every audio transformation.

Source code(tar.gz)
Source code(zip)
0.2.0(Jun 29, 2021)

Removes the Essentia library as a dependency and adds compatibility for multi-channel audio.
Source code(tar.gz)
Source code(zip)
1.0(May 11, 2021)

Source code(tar.gz)
Source code(zip)

Owner

Janne

Music producer, machine learning in MIR & occasional ethical hacker

GitHub Repository

Official implementation of A cappella: Audio-visual Singing VoiceSeparation, from BMVC21

Y-Net Official implementation of A cappella: Audio-visual Singing VoiceSeparation, British Machine Vision Conference 2021 Project page: ipcv.github.io

12 Oct 22, 2022

Neural building blocks for speaker diarization: speech activity detection, speaker change detection, overlapped speech detection, speaker embedding

⚠️ Checkout develop branch to see what is coming in pyannote.audio 2.0: a much smaller and cleaner codebase Python-first API (the good old pyannote-au

2.1k Dec 31, 2022

Audio augmentations library for PyTorch for audio in the time-domain

Related tags

Overview

Audio Augmentations

Usage

Optional

Cite

Comments

Releases(v0.2.3)

v0.2.3(Nov 15, 2021)

0.2.0(Jun 29, 2021)

1.0(May 11, 2021)

Owner

Janne

Official implementation of A cappella: Audio-visual Singing VoiceSeparation, from BMVC21

Neural building blocks for speaker diarization: speech activity detection, speaker change detection, overlapped speech detection, speaker embedding

kapre: Keras Audio Preprocessors

Conferencing Speech Challenge

This is a python package that turns any images into MIDI files that views the same as them

Extract the songs from your osu! libary into proper mp3 form, complete with metadata and album art!

A Python wrapper for the high-quality vocoder "World"

Audio spatialization over WebRTC and JACK Audio Connection Kit

Okaeri-Music is a telegram music bot project, allow you to play music on voice chat group telegram.

Mopidy is an extensible music server written in Python

Learn chords with your MIDI keyboard !

🎵 A repository for manually annotating files to create labeled acoustic datasets for machine learning.

A voice assistant which can be used to interact with your computer and controls your pc operations

Codes for "Efficient Long-Range Attention Network for Image Super-resolution"

Vixtify - Python Controlled Music Player

Delta TTA(Text To Audio) SoftWare

Python implementation of the Short Term Objective Intelligibility measure

This bot can stream audio or video files and urls in telegram voice chats

Bot Music Pintar. Created by Rio

ᴀ ʙᴏᴛ ᴛʜᴀᴛ ᴄᴀɴ ᴘʟᴀʏ ᴍᴜꜱɪᴄ ɪɴ ᴛᴇʟᴇɢʀᴀᴍ ɢʀᴏᴜᴘ ᴏɴ ᴠᴏɪᴄᴇ ᴄᴀʟʟ