High Precision Speech Keyword Spotting Based on Binary Deep Neural Network in FPGA.

Area of Science:

• Speech Processing
• Machine Learning
• Embedded Systems

Background:

• Deep Neural Networks (DNNs) excel in Keyword Spotting (KWS) but are computationally intensive for IoT devices.
• Binary Neural Networks (BNNs) reduce resource needs but often sacrifice accuracy due to information loss during quantization.
• This accuracy drop in BNN-KWS can lead to increased energy consumption from false wake-ups on edge devices.

Purpose of the Study:

• To develop a novel Probability Smoothing Enhanced Binarized Neural Network (PSE-BNN) for efficient KWS on resource-constrained IoT edge devices.
• To achieve a balance between computational complexity and recognition accuracy, enabling deployment on FPGA platforms.
• To mitigate accuracy degradation caused by binarization in KWS systems.

Main Methods:

• Proposed a PSE-BNN model with two modules: preliminary recognition extraction and result recognition.
• The result recognition module uses temporal correlation to denoise features and enhance quantized model performance.
• Leveraged information theory principles to reduce output distribution's conditional entropy via probability smoothing.

Main Results:

• Achieved 97.29% accuracy on the Google Speech Commands Dataset (GSCD).
• On a Xilinx VC707 FPGA, the PSE-BNN used 1939 LUTs, 832 FFs, and 234 Kb storage.
• Improved accuracy by 1.93% and reduced hardware resources by ~65% compared to state-of-the-art BNN-KWS.

Conclusions:

• The PSE-BNN effectively balances computational efficiency and accuracy for KWS on edge IoT devices.
• Probability smoothing successfully reduces noise-induced entropy, enhancing the signal-to-noise ratio.
• The PSE-BNN-FPGA design shows significant potential for deployment on resource-constrained edge IoT applications.