Xiaobin-Rong/gtcrn

What it solves

GTCRN addresses the challenge of performing high-quality speech enhancement (removing noise from audio) on devices with extremely limited computational power. It aims to provide a lightweight alternative to larger models that typically require significant memory and processing resources.

How it works

The project implements a Grouped Temporal Convolutional Recurrent Network. It optimizes efficiency by using a very small number of parameters (48.2K) and low computational operations (33.0 MMACs per second). To maintain streamability for real-time use, it uses an implicit feature rearrangement method via a fully connected layer rather than an explicit shuffle layer.

Who it’s for

This is designed for developers and researchers working on real-time audio processing, specifically those targeting edge devices or low-power hardware where traditional deep learning models are too heavy.

Highlights

• Ultra-low resource usage: Uses only 48.2K parameters and 33.0 MMACs per second.
• High efficiency: Achieves a real-time factor (RTF) of 0.07 on a standard Intel i5 CPU.
• Competitive performance: Outperforms RNNoise and remains competitive with much larger models on VCTK-DEMAND and DNS3 datasets.
• Deployment ready: Includes streaming inference capabilities and support for sherpa-onnx.