| language: en | |
| license: apache-2.0 | |
| tags: | |
| - target-speaker-extraction | |
| - keyword-cue | |
| # DAE-TSE Pretrained Models | |
| ## KCE (Keyword-guided Cue Encoder) | |
| - **File**: `kce/epoch_149.pt` | |
| - **Architecture**: `AEDKWSASRPhone` | |
| - **Training data**: LibriSpeech train-clean-360 + train-other-500 | |
| - **Validation data**: LibriSpeech train-clean-100 | |
| - **Config**: `kce/model.yaml`, `kce/data.yaml` | |
| ## Backbone (DAE-BSRNN) | |
| - **File**: `backbone/models/checkpoint_150.pt` | |
| - **Architecture**: `DAE-BSRNN` (text-aware BSRNN with TC-ASR encoder) | |
| - **Training data**: Libri2Mix-100 mix-clean | |
| - **Config**: `backbone/config.yaml` | |
| > After downloading, update `model_args.tse_model.asr_encoder.checkpoint_dir` in `backbone/config.yaml` to the absolute path of `kce/epoch_149.pt`. | |