Audio Data Augmentation Is All You Need

Data scientists know that data augmentation is an important technique to make the machine learning models they train more robust. While there are loads of resources for data augmentation with image data, it’s difficult to find material for audio augmentation
In this new series, you can learn audio data augmentation from a theoretical and implementation perspectives.
In the first installment of the series,you can learn:
- what’s audio data augmentation
- which use cases it can be fruitfully used for
- the audio data augmentation golden rule
Slides:
github.com/musikalkemist/audi...
=========================
Interested in hiring me as a consultant?
valeriovelardo.com/
Join The Sound Of AI Slack community:
valeriovelardo.com/the-sound-...
Connect with Valerio on Linkedin:
/ valeriovelardo
Follow Valerio on Facebook:
/ thesoundofai
Follow Valerio on Twitter:
/ musikalkemist
=========================
0:00 Intro
0:47 What's data augmentation?
1:38 How does data augmentation work?
3:03 Why audio data augmentation is important
5:23 Use cases for audio data augmentation
7:37 Augmentation chains
8:56 What data should be augmented?
9:41 Offline vs online agumentation
13:23 The golden rule of data augmentation
14:52 Join the community!
15:31 What's next?

[@jeb4148]

Thanks for all the content!! Just finished up the audio signal processing series, excited to jump into this one

[@AbhinayKhoparzi]

This is so insightful. Eagerly waiting for more in the series.

[@CourseWare-xg2wq]

This changed my every view on data augmentation 😮, thnx much Valerio

[@SHADABALAM2002]

Much needed and awaited topic

[@lorenzo87052]

Already waiting next video of this topic! Thank you Valerio!

[@ValerioVelardoTheSoundofAI]

You're welcome Lorenzo!

[@debabratagogoi9038]

Thank you for the nice explanation

[@doyourealise]

thank you so much! because of you this kind of videos exist :)

[@nmirza2013]

I can only say Stay blessed for such a good work

[@ignaziogulino]

Complimenti Valerio!

[@heychamp2054]

Thanks, Valerio for this amazing content!
Question: how can we use onsets for finding similarities between two audio files?
I am currently working on a project. Please help me out.
Any help will be appreciated.

[@kalkidangazahegn5038]

Thanks a lot Valerio, can I apply this audio data augmentation technique to heart sound signal?

[@AbdullahOlcay-wh4or]

Thanks very much Valerio, but why using a coupled code (online augmentation) makes it disadvantageous or advantageous (in offline augmentation). I think this depends on how flexibly one can use augmentation and model together without overfitting and with getting a high accuracy.

[@erikgoron5928]

Great video, thanks Valerio!
Question :
Do you have any tips on how to improve generalizability asides from data augmentation for audio tasks?
I’m currently working on a speech pathology classifier, it works great on single datasets but when I use the trained models on unseen data from different datasets (same task) it doesn’t perform well anymore. Meaning it would not be a good model in production..
Any help is appreciate it.

[@ValerioVelardoTheSoundofAI]

I addressed this topic in a Linkedin post this wekk. Here's the link: www.linkedin.com/posts/valeriovelardo_ml-data-ai-activity-6886601744583860225-ymSw

[@SHADABALAM2002]

Hello Valerio... How can we ensure that semantics are not lost after augmentation.

[@MLDawn]

really nice. well done man!

[@ValerioVelardoTheSoundofAI]

Thank you!

[@maryamkhoshkhoo9030]

Do you have classes to participate for deep learning eeg in python?

[@syllacamara5504]

Thanks a lot for your great work.
Will you do a video on speech recognition with CTC loss function ?
Best regards

[@ValerioVelardoTheSoundofAI]

I'll definitely cover speech recognition in the future. Stay tuned!

[@AdrianFernandezFazio]

Amazing vídeo! Greetings from Argentina!

[@ValerioVelardoTheSoundofAI]

Thanks!

[@madmike0304]

what about audio data augmentation for music generative models?

[@BRAGAARS]

Hello Valerio, just wondering, did you ever work on audio event detection? Been currently trying to build DNNs that detect a specific sound in audio files through image representation such as spectrograms and can't find a lot of things so I am wondering you ever templated doing something similar.

[@ValerioVelardoTheSoundofAI]

I've worked on audio event detection. If the data is good, both spectrograms and Mel spectrograms should be valuable representations to use. If I had to pick one, I'd suggest the latter.

[@Jononor]

Sound Event Detection using Machine Learning (EuroPython 2021)
ua-cam.com/video/JrhsFfCOL-s/v-deo.html

[@mahmodaldahol]

this is exciting topic
thank you

[@ValerioVelardoTheSoundofAI]

Thanks!

[@ont7126]

Hi Valerio! Nice topic, can't wait for your implementation. I am training a CNN with a 7k word audio dataset. Could I reduce underfitting by using data augmentation ?

[@ValerioVelardoTheSoundofAI]

I don't think augmenting data will resolve underfitting. For that, increasing model complexity should help.

[@ont7126]

@@ValerioVelardoTheSoundofAI I am using your CNN implementation for the music genre classification. Underfitting tends to decrease after increasing epochs number. Do you think I must increase model complexity or I should find the ideal number of epochs?

[@ValerioVelardoTheSoundofAI]

@@ont7126 first increase # of epochs. If you still underfit, try increasing complexity.

[@ivanstepanovftw]

Audio in this video augmented too? :)

[@sutarorem8297]

I love you so much

[@Sawaedo]

If there would be a conversion from real_data to augmented_data, how much augmented_data do we need to compensate a decrease in real_data?

[@ValerioVelardoTheSoundofAI]

You should keep the original data in the dataset and add the augmented samples to it. As a general rule I would produce 2-4X the original data. More than that you start having too much redundancy.

[@Bindassmanus]

Sir Can you help for detection of vowel like region in speech signal i have my mtech dissertation project on it

[@ValerioVelardoTheSoundofAI]

I suggest you to ask a question with specific details of the problems you're encountering in The Sound of AI Slack.

[@Bindassmanus]

@@ValerioVelardoTheSoundofAI ok thanks for replying