config.json

{
  "model_type": "demodokos-vocal-spectrography",
  "architecture": "ECAPA-TDNN",
  "version": "1.0.0",
  "encoder": {
    "mel_dim": 128,
    "embedding_dim": 2048,
    "channels": [
      512,
      512,
      512,
      512,
      1536
    ],
    "kernel_sizes": [
      5,
      3,
      3,
      3,
      1
    ],
    "dilations": [
      1,
      2,
      3,
      4,
      1
    ],
    "attention_channels": 128,
    "res2net_scale": 8,
    "se_channels": 128
  },
  "preprocessing": {
    "sample_rate": 24000,
    "n_fft": 1024,
    "num_mels": 128,
    "hop_size": 256,
    "win_size": 1024,
    "fmin": 0,
    "fmax": 12000,
    "center": false,
    "mel_norm": "slaney",
    "window": "hann"
  },
  "training": {
    "objective": "vocal_spectrography",
    "loss": "AAM-Softmax",
    "embedding_normalization": "L2",
    "dtype": "bfloat16"
  },
  "torch_dtype": "bfloat16"
}