Random amplification coefficient as a data augmentation feature

Feature requests
#1

I primarily use Edge Impulse for building models for audio classification that run on the Syntiant TinyML board. I’ve noticed that to achieve better precision, I often need to apply various amplification factors to my raw data (WAV files). As a result, I have various audio samples with different volumes (low, medium, high) for both the detected class and noise. Typically, I perform this task in separate audio software, which is quite time-consuming.
I wonder if this could be integrated into the data augmentation section. I could choose a range of amplification factors (e.g: 0.2 to 10), and my samples would be multiplied by a random number from this range.
Without this step, I usually experience poor classification accuracy, and it appears that the model tends to classify based on the average sample volume rather than the frequency content.

1 Like
#2

Hi @Kroun1

Thanks for suggesting this feature let me log it to our tech team. Anything additional to add here, or some screenshots / further details of the steps you take?:

The request is to introduce a “Random Amplification Coefficient” as a data augmentation feature in Edge Impulse to assist users, particularly those working on audio classification models for devices like the Syntiant TinyML board. This feature aims to streamline the process of applying various amplification factors to raw audio data (e.g., WAV files), enhancing model precision without external software.

Best

Eoin

#3

Hi, thank you for logging this request! I have nothing further to add or share about it at the moment