EigenCodec: Ultra-low bitrate narrowband speech coding via eigen-informed subspace projection and multi-scale...

Narrowband speech coding at ultra-low bitrates remains challenging, primarily due to substantial information loss in highly compressed latent spaces and the computational constraints of lightweight edge models. In this paper, we propose EigenCodec, an efficient neural speech coding framework explicitly optimized for this regime. To enhance representation efficiency, we introduce the multi-scale depthwise aggregation block. By synergizing parallel depthwise convolutions with channel attention mechanisms and lightweight feed-forward networks, this module efficiently aggregates multi-resolution temporal dependencies, reducing parameter redundancy while maintaining receptive field density. Furthermore, addressing the information bottleneck, we propose an eigen-informed subspace projection strategy. By leveraging statistical priors from principal component analysis to initialize the quantization layers, we explicitly guide the encoder to map features into high-variance directions, thereby mitigating information loss during dimensionality reduction. Experimental evaluations on the 8 kHz-sampled LibriSpeech dataset show that EigenCodec yields a competitive balance between computational efficiency and reconstruction fidelity compared to existing baselines, delivering intelligible and natural-sounding narrowband speech reconstruction at 600 and 800 bps.