configs/config.yaml

sample_rate: 24000
model:
  autoencoder:
    _target_: models.autoencoder.waveform.stable_vae.StableVAE
    encoder:
      _target_: models.autoencoder.waveform.stable_vae.OobleckEncoder
      in_channels: 1
      channels: 128
      c_mults:
      - 1
      - 2
      - 4
      - 8
      strides:
      - 2
      - 4
      - 6
      - 10
      latent_dim: 256
      use_snake: true
    decoder:
      _target_: models.autoencoder.waveform.stable_vae.OobleckDecoder
      out_channels: 1
      channels: 128
      c_mults:
      - 1
      - 2
      - 4
      - 8
      strides:
      - 2
      - 4
      - 6
      - 10
      latent_dim: 128
      use_snake: true
      final_tanh: false
    io_channels: 1
    latent_dim: 128
    downsampling_ratio: 480
    sample_rate: 24000
    pretrained_ckpt: ckpts/1m.pt
    bottleneck:
      _target_: models.autoencoder.waveform.stable_vae.VAEBottleneck
  backbone:
    _target_: models.dit.mask_dit.UDiT
    img_size: 500
    patch_size: 1
    in_chans: 128
    out_chans: 128
    input_type: 1d
    embed_dim: 1024
    depth: 24
    num_heads: 16
    mlp_ratio: 4.0
    qkv_bias: false
    qk_scale: null
    qk_norm: layernorm
    norm_layer: layernorm
    act_layer: geglu
    context_norm: true
    use_checkpoint: true
    time_fusion: ada_sola_bias
    ada_sola_rank: 32
    ada_sola_alpha: 32
    cls_dim: null
    context_dim: 1024
    context_fusion: cross
    context_max_length: null
    context_pe_method: none
    pe_method: none
    rope_mode: shared
    use_conv: true
    skip: true
    skip_norm: true
  cfg_drop_ratio: 0.2
  _target_: models.flow_matching.SingleTaskCrossAttentionAudioFlowMatching
  content_encoder:
    _target_: models.content_encoder.content_encoder.ContentEncoder
    embed_dim: 1024
    text_encoder: None
    speech_encoder:
      _target_: models.content_encoder.star_encoder.star_encoder.QformerBridgeNet
      load_from_pretrained: ckpts/exp0_best.pt
  pretrained_ckpt: model.safetensors