Underwater Embedded Offline Speech Recognition System with Ocean Noise Augmentation
Fenghao Jin, Ruiqin Zhao, Wei Zhang · 2024
The embedded offline speech recognition system deploys a pre-trained end-to-end model on an embedded device. It maintains high accuracy while eliminating reliance on network connectivity and reducing the impact of network instability. In this paper, to address the impact of noise in marine environments on speech recognition accuracy, we augment the Aishell-1 dataset with an ocean noise dataset to provide a more realistic training environment for the model. The Transducer model, which is an end-to-end model, is trained using the augmented data and subsequently quantized for deployment on a Raspberry Pi 4B. By evaluating the Character Error Rate (CER) and Real-Time Factor(RTF) of the system under different configurations and parameter settings, we propose a solution for building an underwater embedded offline speech recognition system.