Pruning for Enhancing Model Efficiency in Single-Channel Speech Separation
- Status
- Finished
- Type
- Master Project
- Announcement date
- 29 Jul 2025
- Student
- Daniel Zirat
- Mentors
- Research Areas
Pruning techniques enables to develop more compact and efficient machine learning models. This not only reduces the demand on resources but also enhances performance across a broad spectrum of applications. Thus, model pruning is fundamental to realizing powerful AI systems that prioritize both efficiency and effectiveness, extending from mobile
applications to edge computing environments. This thesis advocates the adoption of a model pruning strategy to achieve the goal of developing more efficient and streamlined models for single-channel source separation. Specifically, parameterized pruning applied on two different models, WaveSplit and ConvTasNet is investigated to navigate the previously mentioned constraints. In the experiments, the ConvTasNet and WaveSplit models were specifically pruned using strategies (epoch-wise, periodic, and gradual) to evaluate the impact on model size, efficiency, and separation performance. The goal was to analyze the trade-off between model complexity and performance in resource-constrained environments such as mobile or embedded systems. The results clearly demonstrate that model complexity can be significantly reduced through pruning techniques without substantially compromising separation quality. On average, ConvTasNet was able to reduce its number of weights by 38% without impairing two-speaker separation. For Wavesplit, an average of 24% of the weights in the convolutional block could be removed without degrading source separation.