TurboCRNN: A Robust and Efficient Architecture for Wake Word Detection

Pitch video

Watch here

Introduction

This repository contains the implementation of TurboCRNN, a novel and optimized wake word detection model designed to achieve minimal false predictions, reduced latency, and low complexity. The project focuses on preprocessing audio data, extracting features, and using advanced neural network architectures to accurately detect wake words.

Motivation

Speech is a natural and efficient way of interacting with devices, enhancing multitasking and user experience. Wake word detection is crucial for activating voice assistants and smart systems. Despite significant research, there is a need for lightweight models with reduced false positives that can run efficiently on low-power systems.

Problem to solve

To build an optimized wake word detection model with minimal false predictions, reduced latency, and low complexity, capable of running efficiently on low-power systems such as mobile devices.

Dataset

• Dataset Used: Google Speech Commands Dataset
• Details:
  • 30 Words in total
  • 65,000 Audio Samples in total
  • .wav Format
  • Almost 1 second in duration
• Wake Word: Marvin
• Dataset Size:
  • Marvin: 1746 Samples
  • False Data: 1769 Samples

Audio Preprocessing

Denoising

• Technique: Wavelet Decomposition using db4 wavelets from the Daubechies family.
• Process:
  • Signal is broken down into frequency bands and noise is removed using thresholding.
  • Denoised audio is reconstructed using inverse wavelet transform.

Data Augmentation

• Techniques Implemented:
  • Time Scaling
  • Pitch Scaling
  • White Noise Addition
  • Polarity Inversion
  • Random Gain
• Dataset Size After Augmentation:
  • True Samples: 10476
  • False Samples: 10614
  • Total: 21090

Feature Extraction

Techniques

• Mel Frequency Cepstral Coefficients (MFCCs): Commonly used feature extraction technique in speech recognition.
• Linear Frequency Cepstral Coefficients (LFCCs): Compared against MFCCs in our research for better performance.

Performance Comparison

Model	MFCC Accuracy	LFCC Accuracy
DNN	0.7141	0.8113
CNN	0.9374	0.9317
LSTM	0.9075	0.9692
GRU	0.9431	0.9787
CRNN_GRU_LSTM	0.9493	0.9815
CRNN_GRU_GRU	0.9663	0.982

Neural Networks

Baseline Model Architectures

• DNN
• CNN
• LSTM
• GRU

Proposed Model Architectures

• CRNN_GRU_GRU: Convolutional Layers + GRU Layer + GRU Layer
• CRNN_GRU_LSTM: Convolutional Layers + GRU Layer + LSTM Layer

Experimentation

• Multi-head Attention
• L2 Regularizer
• Extra Dense Layer

Optimizer & Loss Functions

• Optimizer: AMS Grad Optimizer
• Loss Function: Categorical Cross Entropy
• Train-Test Split: 80% Training, 10% Validation, 10% Testing
• Training: 120 Epochs, Learning Rate = 0.01

Model Performance Comparison

Model	Accuracy	False Positives	False Negatives
DNN	0.8113	338	60
CNN	0.9317	73	71
LSTM	0.9692	47	18
GRU	0.9787	25	20
CRNN_GRU_LSTM	0.9815	18	21
CRNN_GRU_GRU	0.982	18	20

Setup Instructions

1. Clone the Repository

git clone https://github.com/dhairya1702/TurboCRNN
cd TurboCRNN

2. Set Up Python Environment

# Using virtualenv
pip install virtualenv
virtualenv venv
source venv/bin/activate  # On Windows, use venv\Scripts\activate

# Using conda
conda create --name myenv python=3.8
conda activate myenv

3. Install Required Libraries

pip install numpy scipy librosa matplotlib jupyter

4. Running Jupyter Notebooks

# Install Jupyter Notebook
pip install jupyter

# Start Jupyter Notebook
jupyter notebook

5. Running Scripts and Notebooks

Notebooks:

MFCC_Extraction_Augmented.ipynb
LFCC_Extraction_Augmented.ipynb
Feature Extraction Techniques - Testbed.ipynb
Data Denoising.ipynb
Data Denoising - TestBed.ipynb
Data Augmentation.ipynb

Open these notebooks in Jupyter and run each cell sequentially.

python Dataset Creation Script.py

6. GRU LSTM Models

GRU LSTM 2.ipynb
GRU LSTM 5.ipynb
GRU LSTM 8.ipynb

Open these notebooks in Jupyter and run each cell sequentially.

Performance Summary

Best Performer in 100k Parameter Budget: CRNN_GRU_LSTM Model Variant #5

With Extra Dense Layer, Without Attention, Without Regularizer - TurboCRNN 75% and 91% reduction in false predictions when compared to CNN and DNN baselines. 74ms and 80ms latencies in the presence and absence of GPU. 87k model parameters, low model complexity.