BASENet: Band-Adapted Speech Enhancement Network with Cross-Band Attention

2026-06-10 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Speech Processing · Depth: Expert, quick

Summary

BASENet is a novel speech enhancement network designed with a frequency-adapted architecture that addresses the non-uniform spectral resolution of human hearing. It partitions the audio spectrum into Bark-scale bands, assigning scaled-capacity encoders based on critical-band density, which results in deeper processing for perceptually dense low frequencies and lighter processing for high frequencies. The network incorporates a cross-band attention module to capture harmonic dependencies across bands using compact frequency-pooled representations, operating with linear complexity. Built upon inverted residual blocks, dense connectivity, and a convolutional recurrent network, BASENet achieves a 3.55 PESQ score and STOI~96% on the VoiceBank+DEMAND dataset. Notably, it uses only 0.83M parameters and 7.3 G MACs, making it the most parameter-efficient method among those exceeding 3.50 PESQ. A causal variant, achieving 3.44 PESQ, further demonstrates its suitability for real-time streaming on resource-constrained devices.

Key takeaway

For Machine Learning Engineers developing real-time speech enhancement systems, BASENet offers a highly efficient architecture. You should consider its frequency-adapted design and cross-band attention for optimizing performance on resource-constrained devices. Its causal variant, achieving 3.44 PESQ with only 0.83M parameters, provides a strong benchmark for balancing high-quality output with minimal computational overhead, enabling deployment in edge computing scenarios.

Key insights

BASENet enhances speech by adapting network capacity to Bark-scale frequency bands and using cross-band attention for efficient, perceptually-aware processing.

Principles

• Human hearing has non-uniform spectral resolution.
• Adapt network capacity to perceptual frequency density.
• Cross-band attention captures harmonic dependencies.

Method

BASENet partitions the spectrum into Bark-scale bands, assigning scaled-capacity encoders. It uses a cross-band attention module with frequency-pooled representations and is built on inverted residual blocks and a convolutional recurrent network.

In practice

• Deploy causal variant for real-time streaming.
• Optimize speech enhancement for resource-constrained devices.
• Improve PESQ/STOI with minimal parameters.

Topics

• Speech Enhancement
• Neural Networks
• Bark Scale
• Cross-Band Attention
• Real-time Audio Processing
• Resource-Constrained Devices

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• GetNetUPAM: Ecologically Informed Nested Cross-Validation and Noise-Robust Attention for Marine Bioacoustic Monitoring — Ecologically informed nested cross-validation and adaptive attention improve marine bioacoustic model stability and generalization.
• From DSP to AI: Evolving Approaches in Audio Processing — Hybrid audio AI systems combine DSP's physical grounding with ML's adaptability for robust, perceptually aligned performance.
• ART: A Faster Way to Train Speech Recognition Models — ART accelerates RNN-T training by combining CTC for baseline alignment with residual learning for refinement.
• RAIL: Rethinking Auditory Intelligence in Large Audio-Language Models with a CHC-Grounded Benchmark — LALM evaluation should shift from task-centric metrics to human-centric cognitive frameworks like CHC for deeper auditory intelligence assessment.
• Lung-SRAD: Spectral-Aware Regularized Audio DASS with Dual-Axis Patch-Mix Contrastive Learning for Respiratory Sound Classification — Lung-SRAD uses State Space Models and spectral-aware regularization to improve respiratory sound classification by preserving high-frequency details.
• In Silico Modeling of the RAMPHO Buffer: Dissociating Informational and Energetic Masking via Phonetic Entropy in Deep Neural Networks — Modeling phonetic entropy in DNNs can dissociate informational from energetic masking in multi-talker environments.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.