This repository is the official implementation of paper "GTCRN: A Speech Enhancement Model Requiring Ultralow Computational Resources". The paper has been accepted by ICASSP 2024, and will be accessible for online download starting from March 15th or later, according to the official schedule.
Audio examples are available at Audio examples of GTCRN.
Grouped Temporal Convolutional Recurrent Network (GTCRN) is a speech enhancement model requiring ultralow computational resources, featuring only 23.7 K parameters and 39.6 MMACs per second. Experimental results show that our proposed model not only surpasses RNNoise, a typical lightweight model with similar computational burden, but also achieves competitive performance when compared to recent baseline models with significantly higher computational resources requirements.
Pre-trained models are provided in checkpoints folder, which were trained on DNS3 and VCTK-DEMAND datasets, respectively.
The inference procedure is presented in infer.py.
A streaming GTCRN is provided in stream folder, which demonstrates an impressive real-time factor (RTF) of 0.07 on the 12th Gen Intel(R) Core(TM) i5-12400 CPU running at 2.50 GHz.
SEtrain: A training code template for DNN-based speech enhancement.
TRT-SE: An example of how to convert a speech enhancement model into a streaming format and deploy it using ONNX or TensorRT.