A Simplified Post-Training Quantization Method for Multi-Channel Speech Separation Model
Xu Zhang, Changchun Bao, Xue Yang · 2024
In recent years, the performance of speech separation has been greatly improved by using neural networks. However, the increasing complexity of neural network design usually requires a large amount of computing resources and storage space, which limits its application in resource-constrained environments. To address this issue, the model compression techniques have emerged. By quantizing the weights and activation values of the model, the computational cost can be significantly reduced while maintaining separation performance. This paper introduces a simplified post-training quantization (PTQ) method to compress existing speech separation models by using 8-bit precision for both weight and activation instead of 32-bit precision. The experiments on the spatialized version of the WSJ0-2mix corpus show that the proposed method keeps comparable performance with the baseline speech separation systems in terms of SDR scores, while moderately reduces the complexity.