gsu-dbnet-snn-speech-enhancement

name	gsu-dbnet-snn-speech-enhancement
description	GSU-DBNet: Dual-Branch Spiking Neural Network for neuromorphic speech enhancement with Gated Spiking Units (GSU). Achieves PESQ 3.04 with only 394K params (4.5-10.6% of ANN models). (arXiv: 2606.23761)
tags	["SNN","speech-enhancement","neuromorphic","gated-spiking-unit","dual-branch","energy-efficient","Interspeech2026"]

GSU-DBNet: Dual-Branch SNN for Speech Enhancement

Paper Reference

• Title: Neuromorphic Speech Enhancement with Dual-Branch Spiking Neural Networks
• arXiv: 2606.23761
• Authors: Taiyu Meng, Wenbin Jiang, Haoyi Zhang, Yuhan Zhou, Haibing Yin
• Submitted: June 22, 2026
• Venue: Submitted to Interspeech 2026
• Categories: cs.SD, cs.AI, eess.AS

Core Methodology

Architecture: GSU-DBNet

Novel dual-branch spiking neural network with Gated Spiking Units (GSU) for speech enhancement.

Key Components

1. Gated Spiking Unit (GSU): Novel neuron model with gating mechanism addressing binary activation limitations
2. Dual-Branch Architecture: Simultaneously models:
   • Speech magnitude spectrum → magnitude spectral mask
   • Complex spectrum → complex spectral mask
3. Dual-Path GSU Module: Exploits both temporal and frequency dimensions for enhanced spatiotemporal feature representation
4. Spike-Driven Processing: Entire pipeline uses event-driven computation

Innovation Points

• First dual-branch SNN for speech enhancement
• GSU gating mechanism overcomes binary activation limitations of traditional SNNs
• Simultaneous magnitude and complex spectrum processing (bio-inspired: mimics cochlear dual processing)
• Extreme parameter efficiency: 394K params

Results

• PESQ Score: 3.04 (state-of-the-art among SNN-based methods)
• Parameter Count: Only 394K parameters
  • 4.5%--10.6% of representative ANN-based models
• Energy Efficiency: Inherent spike-driven computation enables ultra-low power deployment

Neuroscience Connection

• Dual-branch processing mirrors auditory system's parallel magnitude/phase processing
• GSU gating resembles cortical inhibitory gating mechanisms
• Temporal-frequency dual-path mimics auditory cortex feature hierarchy
• Demonstrates neuromorphic computing's viability for real-world audio tasks

Practical Applications

• Hearing aids and cochlear implants
• Voice assistants on edge devices
• Real-time noise cancellation
• Telecommunications
• IoT audio processing

Activation Keywords

• GSU-DBNet
• gated spiking unit
• dual-branch SNN
• neuromorphic speech enhancement
• SNN speech processing
• Interspeech 2026 SNN
• energy-efficient speech

Related Skills

• [[spiketimer-snn-copyright-protection]] - SNN temporal coding
• [[emrformer-neuromorphic-amr]] - Neuromorphic signal processing
• [[edgespike-edge-iot-snn]] - Edge SNN deployment
• [[neuromorphic-supremacy]] - Neuromorphic computing framework