IncepTCN: A Multi-Scale Inception-Temporal Architecture for Contextual Feature Learning in Raw Speech Enhancement

Abstract Speech enhancement improves intelligibility and quality of noisy signals but remains challenging at low signal to noise ratio (SNR) and under real-time constraints. Time-domain approaches perform end-to-end waveform modeling, with many recent methods achieving multi-scale feature extraction. Most rely on dilation variation, temporal-only scaling, or sequential multi-scale modules with attention and gating, which improve accuracy but increase complexity and hinder real-time use. Therefore, our contribution is an inception temporal convolutional neural network called IncepTCN, a lightweight encoder–decoder architecture that integrates a parallel multi-kernel inception module into a causal dilated TCNN to capture fine-grained speech details and broader context. By using different convolutional sizes within a single module, the model sees speech at multiple scales simultaneously, improving separation from complex, non-stationary noise. Experiments on the Microsoft Scalable Noisy Speech Dataset (MS-SNSD) show that the proposed IncepTCN model outperforms conventional TCNs, achieving improvements of up to 0.53 in the Perceptual Evaluation of Speech Quality (PESQ), 9.14% in the Short-Time Objective Intelligibility (STOI), and 24.1 dB in the Segmental Signal-to-Noise Ratio (SSNR), while generalizing effectively to unseen noise types and low-SNR conditions. Furthermore, the proposed IncepTCN maintains computational efficiency with 5.754M trainable parameters and supports real-time inference with low computational overhead.