NRSC-KG: Noise-Robust Speech Comprehension Based on Knowledge Graph for HRI 

Abstract

Speech comprehension is the key to Human-Robot Interaction (HRI), however, noise energetic masking can greatly reduce the performance of speech comprehension. Simply using Speech Enhancement (SE) methods can easily lead to suboptimal solutions. To alleviate these issues, this paper proposes a Noise-Robust Speech Comprehension system based on Knowledge Graph (NRSC-KG), which designs robust methods in Automatic Speech Recognition (ASR) and ASR error correction. In ASR, we propose a Self-adaptive Speech Feature Fusion Network (SSFF-Net) to dynamically fuse enhanced speech features with the original speech features. In ASR error correction, we propose a divergent apperceive error correction method based on domain-specific Knowledge Graph (KG). The experiment on AISHELL-1 speech corpus shows that compared to the traditional joint training of SE and ASR, our method achieves a relative Character Error Rate (CER) reduction of 12.7%. We validated the KG based error correction method on our customized Herbal Cuisine ordering Dialogue speech dataset (HCDialogue), and the results showed that our method achieved a relative CER reduction of 5.0% and 2.1%, respectively, compared to the language model based and end-to-end neural network based methods. It’s indicated that the KG based method has great potential in improving speech comprehension performance.