A Lightweight End-to-End Speech Recognition System on Embedded Devices

In industry, automatic speech recognition has come to be a competitive feature for embedded products with poor hardware resources. this work, we propose tiny end-to-end model that is lightweight and easily deployable on edge platforms. First, instead of sophisticated network structures, such as recurrent neural networks, transformers, etc., the mainly uses convolutional networks its backbone. This ensures our supported by most software development kits devices. Second, adopt basic unit MobileNet-v3, which performs well in computer vision tasks, integrate features hidden layer at different scales, thus compressing number parameters less than 1 M achieving an accuracy greater some traditional models. Third, order further reduce CPU computation, directly extract acoustic representations from 1-dimensional waveforms use self-supervised learning approach encourage convergence model. Finally, solve problems where resources are relatively weak, prefix beam search decoder dynamically extend path optimized pruning strategy additional initialism language capture probability between-words advance avoid premature correct words. experiments, according evaluation categories, outperformed several models used devices related work.