A Neural Codec for Bone-Conducted Speech With Joint Denoising and Bandwidth Extension

In Internet of Things (IoT) communication scenarios such as wearable and industrial systems operating in extremely noisy environments, Bone-Conducted Microphone (BCM) speech inevitably suffers from noise interference. To address this issue, a neural codec for bone-conducted speech with joint denoising and bandwidth extension is proposed in this paper. Speciffcally, an encoder–decoder architecture composed of the MP-SENet and the BigVGAN vocoder is built as the backbone. On this basis, a Forward Time-Frequency Transformer (FTM) is introduced to effectively capture both local and global time-frequency dependencies in the speech signal. To improve the capacity of speech signal modeling, a Backward Time-Frequency Transformer (BTM) is presented to model the information flow in reverse, complementing the forward modeling and enabling more comprehensive extraction of time-frequency features. Subsequently, KalmanNet is designed to enhance the ability to capture temporal dependencies and dynamic information, compensating for the transformer’s limited effectiveness in modeling continuous dynamics. Finally, a Dual-Path Magnitude-Phase (DPMP) module is proposed to feature an interactive dual-stream architecture where magnitude and phase streams communicate with each other to supplement high-frequency components of the speech signal. The proposed method enables simultaneous speech encoding and enhancement under different SNRs using a single network. Furthermore, it supports coding at different bitrates without requiring modifications to the network architecture or retraining. Simulation experiments demonstrate the feasibility of the proposed approach, highlighting its potential for robust and efficient speech transmission in IoT-enabled communication.