FastWave: A Hardware Architecture for Audio Neural Networks

When Siri, Alexa, Cortana, Google Assistant or your other favorite digital assistant talk to you, they rely on neural networks to create the audio file that speaks to you. WaveNet is a deep neural network for generating audio that provides amazingly accurate results. Yet, this process is slow and cannot be performed in real-time. Our FastWave hardware architecture accelerates this process providing a 10x decrease in the time required to generate the audio file as compared to a state of the art GPU solution. This is the first hardware accelerated platform for autoregressive convolutional neural networks.

FastWave is being presented at the International Conference on Computer-aided Design (ICCAD). ICCAD is one of the top conferences for topics related to hardware design automation. The paper was developed as a project in my CSE 237C class, which teaches hardware design and prototyping using high level synthesis. Shehzeen Hussain, Mojan Javaheripi, and Paarth Neekhara developed the initial idea as a final class project. They continued their work after class and the end result is the paper, FastWave: Accelerating Autoregressive Convolutional Neural Networks on FPGA.