Dynamic convolution models for cross-frontend keyword spotting.

Area of Science:

• Speech processing
• Machine learning
• Artificial intelligence

Background:

• Keyword spotting (KWS) systems are crucial for voice-activated devices.
• Existing KWS methods face challenges with acoustic variability and noisy environments.
• Developing efficient and robust KWS models is an ongoing research area.

Purpose of the Study:

• To propose a novel and efficient keyword spotting method.
• To enhance model generalization and robustness across diverse acoustic conditions.
• To achieve high accuracy with minimal computational resources.

Main Methods:

• Integration of a dynamic convolution model for adaptive acoustic pattern capture.
• Implementation of a cross-frontend mutual learning strategy to leverage complementary features.
• Utilizing the Google Speech Commands dataset for experimental validation.

Main Results:

• Achieved 97% accuracy on the Google Speech Commands dataset.
• The model requires only 62K parameters and 6.11M FLOPs, demonstrating high efficiency.
• Exhibited strong robustness in noisy environments, maintaining performance under low signal-to-noise ratios.

Conclusions:

• The proposed keyword spotting method is highly efficient and accurate.
• The dynamic convolution and mutual learning strategies contribute to robust performance.
• This approach presents a promising solution for real-world keyword spotting applications.