scripts/4_train/voicebank-demand/configs/speech-noise,unet.yaml

---
task_name: voicebank-demand
train:
    input_source_types:
        - speech
        - noise
    target_source_types: 
        - speech
    indexes_dict: "indexes/voicebank-demand/sr=44100_chn=1/train/speech-noise.pkl"
    sample_rate: 44100
    channels: 1
    segment_seconds: 3.0
    model_type: UNet
    loss_type: l1_wav
    optimizer_type: Adam
    mini_data: False
    augmentations:
        mixaudio:
            speech: 1
            noise: 1
        pitch_shift:
            speech: 0
            noise: 0
        magnitude_scale:
            speech:
                lower_db: 0
                higher_db: 0
            noise:
                lower_db: 0
                higher_db: 0
        swap_channel:
            speech: False
            noise: False
        flip_axis:
            speech: False
            noise: False
    batch_data_preprocessor: BasicBatchDataPreprocessor
    evaluation_callback: Default
    learning_rate: 1e-3
    batch_size: 16
    precision: 32
    steps_per_epoch: 10000  # Every 10000 steps is called an `epoch`.
    evaluate_step_frequency: 10000     # Evaluate every #evaluate_step_frequency steps.
    save_step_frequency: 50000  # Save every #save_step_frequency steps.
    early_stop_steps: 500001
    warm_up_steps: 1000
    reduce_lr_steps: 15000
    random_seed: 1234
    resume_checkpoint: ""

evaluate:
    segment_seconds: 30.0
    batch_size: 1