navi2311 /  Efficient_Keyword_Spotting_Accelerator_for_Caravel

Created
Maintained by navi2311
"Efficient Keyword Spotting Accelerator for Caravel SoC: A CNN-Based Approach" is a project aimed at designing a hardware accelerator optimized for keyword spotting tasks on the Caravel System-on-Chip (SoC) platform. Leveraging Convolutional Neural Networks (CNNs), the accelerator aims to achieve high accuracy and low-latency performance while minimizing power consumption. The project focuses on developing a lightweight CNN architecture tailored for embedded deployment, integrating it into the Caravel SoC environment, and providing a scalable and energy-efficient solution for keyword spotting applications in resource-constrained devices.
Members 1
Naveen committed 8 months ago

Project Proposal: Efficient_Keyword_Spotting_Accelerator_for_Caravel

Introduction

The rapid advancement of Keyword Spotting (KWS) technology has revolutionized user interactions with devices through voice commands. This project proposal aims to leverage the power of generative AI to develop an open-source hardware accelerator designed explicitly for KWS applications on the Caravel System-on-Chip (SoC). By integrating a CNN-based KWS model with generative AI optimization techniques, we aim to create an energy-efficient and high-performance KWS accelerator that seamlessly integrates into the Caravel SoC environment.

Objectives

  • Optimize KWS Model: Utilize generative AI to optimize the CNN-based KWS machine learning model for low latency and high accuracy.
  • Enhance Audio Feature Extraction: Apply generative AI techniques to optimize audio feature extraction methods (e.g., MFCC) for improved KWS performance.
  • Design Hardware Accelerator: Implement the optimized KWS model and audio feature extractor as a hardware accelerator using Efabless’ chipIgnite and Caravel SoC.

Methodology

Data Collection and Preprocessing

  • Dataset: Gather a labeled dataset of audio samples for the keywords "Play", "Pause", "Skip", "Volume Up", and "Volume Down".
  • Data Preprocessing: Extract MFCC features and apply data augmentation techniques like time stretching and adding background noise.

CNN Architecture and Training

  • CNN Architecture: Design a CNN architecture suitable for KWS using frameworks like TensorFlow with the help of chatgpt.
  • Training: Train the CNN model on the preprocessed dataset using categorical cross-entropy loss and Adam optimizer with regularization techniques.

Generative AI Optimization

  • Model Optimization: Utilize generative AI tools such as chatGPT or Copilot to optimize the KWS model architecture and parameters.
  • Feature Extraction Optimization: Apply generative AI techniques to optimize the MFCC feature extraction process.

Hardware Implementation

  • chipIgnite and Caravel SoC: Implement the optimized KWS model and feature extractor as a hardware accelerator in Caravel SoC.
  • Optimization: Apply model pruning and quantization techniques to reduce the model size and improve energy efficiency.