Leveraging Pruning & Quantization for Real-Time Audio Applications – Dharanipathi Rathna Kumar

Join Us For ADC24 - Bristol - 11-13 November 2024
More Info: https://audio.dev/
@audiodevcon​

Leveraging Pruning and Quantization for Efficient Real-Time Audio Applications - Dharanipathi Rathna Kumar - ADCx India 2024

In the constantly changing world of audio processing, the demand for real-time response and high-quality output is a relentless challenge, especially in the presence of computational constraints. The rapid growth in model complexity, especially in deep learning, has made it challenging to deploy complex models involving architectures such as TCN, LSTM, etc. on resource-constrained devices and/or to achieve fast inference speeds for tasks such as real-time audio effects, audio style transfer, and source separation. Model compression is vital to address these challenges, making it possible to retain high performance while using fewer computational resources. This talk delves deep into two paramount model compression techniques, namely pruning, and quantization, and explores their applicability in the context of audio applications.

Pruning is a method of eliminating redundant or less contributive weights from the model to reduce the computation resources required to run the network. We'll explore its variants, methodologies, and outcomes, and how it can drastically reduce computational complexity without significantly undermining model performance. Quantization is the process of reducing the precision of the weights, biases, and activations such that they consume less memory. By reducing the bit-width of model parameters, we can achieve sizeable savings in memory and computational power, making it indispensable for on-device audio applications for real-time audio contexts.

In this presentation, I elucidate that by adopting strategic weight pruning and parameter quantization, it is feasible to significantly enhance the efficiency of sophisticated audio models, paving the way for robust, real-time audio processing even in resource-constrained environments.

Link to Slides: https://data.audio.dev/talks/ADCxIndia/2024/pruning-and-quantization-for-efficient-real-time-audio-applications.pdf
_

Edited by Digital Medium Ltd - online.digital-medium.co.uk
_

Organized and produced by JUCE: https://juce.com/
_

Special thanks to the ADC24 Team:

Sophie Carus
Derek Heimlich
Andrew Kirk
Bobby Lombardi
Tom Poole
Ralph Richbourg
Prashant Mishra

#adc #audiodev #dsp #audio #audio