model_config.json

{
    "summary": {
        "quantizer_dim": 3584,
        "codebook_per_group_per_residual": 3584,
        "group": 1,
        "residual": 1,
        "original_residual_dim": 1024,
        "codebook_upsample": 3.5,
        "codebook_dim": 3584
    },
    "base_model": "QWen2.5-7B-Pretrain",
    "token_id_offset": 152064,
    "spec_transform": {
        "sampling_rate": 24000,
        "segment_size": 72000,
        "num_mels": 128,
        "n_fft": 1024,
        "hop_size": 256,
        "win_size": 1024,
        "fmin": 0,
        "fmax": 12000,
        "fmax_loss": null
    },
    "encoder": {
        "input_channels": 128,
        "depths": [
            3,
            3,
            9,
            3
        ],
        "dims": [
            256,
            512,
            768,
            1024
        ],
        "drop_path_rate": 0.2,
        "kernel_size": 7
    },
    "decoder": {
        "hop_length": 256,
        "upsample_rates": [
            8,
            4,
            2,
            2,
            2
        ],
        "upsample_kernel_sizes": [
            16,
            12,
            4,
            4,
            4
        ],
        "resblock_kernel_sizes": [
            3,
            7,
            11
        ],
        "resblock_dilation_sizes": [
            [
                1,
                3,
                5
            ],
            [
                1,
                3,
                5
            ],
            [
                1,
                3,
                5
            ]
        ],
        "num_mels": 1024,
        "upsample_initial_channel": 1024,
        "use_template": false,
        "pre_conv_kernel_size": 13,
        "post_conv_kernel_size": 13
    },
    "quantizer": {
        "quantizer_type": "grvq",
        "input_dim": 1024,
        "n_groups": 1,
        "n_codebooks": 1,
        "codebook_size": 32768,
        "codebook_dim": 3584,
        "levels": [
            8,
            5,
            5,
            5
        ],
        "downsample_factor": [
            1
        ],
        "ema_decay": 0.8,
        "codebook_diversity_loss_weight": 1.0,
        "codebook_diversity_temperature": 100.0
    },
    "teacher_quantizer": {
        "quantizer_type": "grvq",
        "input_dim": 1024,
        "n_groups": 2,
        "n_codebooks": 1,
        "codebook_size": 32768,
        "codebook_dim": 3584,
        "levels": [
            8,
            5,
            5,
            5
        ],
        "downsample_factor": [
            2
        ],
        "ema_decay": 0.8,
        "codebook_diversity_loss_weight": 1.0,
        "codebook_diversity_temperature": 100.0
    },
    "descriminators": {
        "MultiPeriodDiscriminator": {
            "periods": [
                5,
                8,
                13,
                19,
                30
            ],
            "kernal_size": 5,
            "stride": 3
        },
        "MultiScaleDiscriminator": {
            "avg_poolings": {
                "kernal_sizes": [
                    6,
                    6
                ],
                "stridess": [
                    3,
                    3
                ],
                "paddings": [
                    3,
                    3
                ]
            },
            "DiscriminatorS": {
                "kernal_sizes": [
                    21,
                    61,
                    61,
                    61,
                    61,
                    61,
                    7
                ],
                "strides": [
                    1,
                    3,
                    3,
                    6,
                    6,
                    1,
                    1
                ],
                "paddings": [
                    10,
                    30,
                    30,
                    30,
                    30,
                    30,
                    3
                ]
            }
        },
        "MultiScaleSTFTDiscriminator": {
            "n_ffts": [
                1024,
                2048,
                512,
                256,
                128
            ],
            "hop_lengths": [
                256,
                512,
                128,
                64,
                32
            ],
            "win_lengths": [
                1024,
                2048,
                512,
                256,
                128
            ],
            "filters": 32,
            "in_channels": 1,
            "out_channels": 1
        }
    }
}