Dynamic Language Group-Based MoE: Enhancing Code-Switching Speech Recognition with Hierarchical Routing
- The configuration file for the experiment is located in
./conf. - The package details for the experimental conda environment are listed in
./environment_packages.txt. - The source code for this experiment can be found in
./src/DLG-MoE. - The logs are available in
./exp.
We implement a highly flexible MoE model, based on the proposed hierarchical routing and dynamic language expert groups, which allows us to flexibly carry out the design of the expert group according to the actual needs and to choose different topk for inference in order to realize the trade-off between performance and speed. And since we are based on the U2++ architecture, we also support streaming inference with different chunksizes.
You just need to prepare the dataset and place it in the ./data and run the following command to reproduce our experiment.
bash train.sh
bash infer.sh
Our code is primarily modified from wenet version 2.0