config/comosvc.json

{
    "base_config": "config/svc/base.json",
    "model_type": "DiffComoSVC",
    "task_type": "svc",
    "preprocess": {
        // data augmentations
        "use_pitch_shift": false,
        "use_formant_shift": false,
        "use_time_stretch": false,
        "use_equalizer": false,
        // acoustic features
        "extract_mel": true,
        "mel_min_max_norm": true,
        "extract_pitch": true,
        "pitch_extractor": "parselmouth",
        "extract_uv": true,
        "extract_energy": true,
        // content features
        "extract_whisper_feature": false,
        "whisper_sample_rate": 16000,
        "extract_contentvec_feature": false,
        "contentvec_sample_rate": 16000,
        "extract_wenet_feature": false,
        "wenet_sample_rate": 16000,
        "extract_mert_feature": false,
        "mert_sample_rate": 16000,
        // Default config for whisper
        "whisper_frameshift": 0.01,
        "whisper_downsample_rate": 2,
        // Default config for content vector
        "contentvec_frameshift": 0.02,
        // Default config for mert
        "mert_model": "m-a-p/MERT-v1-330M",
        "mert_feature_layer": -1,
        "mert_hop_size": 320,
        // 24k
        "mert_frameshit": 0.01333,
        // 10ms
        "wenet_frameshift": 0.01,
        // wenetspeech is 4, gigaspeech is 6
        "wenet_downsample_rate": 4,
        // Default config
        "n_mel": 100,
        "win_size": 1024,
        // todo
        "hop_size": 256,
        "sample_rate": 24000,
        "n_fft": 1024,
        // todo
        "fmin": 0,
        "fmax": 12000,
        // todo
        "f0_min": 50,
        // ~C2
        "f0_max": 1100,
        //1100,    // ~C6(1100), ~G5(800)
        "pitch_bin": 256,
        "pitch_max": 1100.0,
        "pitch_min": 50.0,
        "is_label": true,
        "is_mu_law": true,
        "bits": 8,
        "mel_min_max_stats_dir": "mel_min_max_stats",
        "whisper_dir": "whisper",
        "contentvec_dir": "contentvec",
        "wenet_dir": "wenet",
        "mert_dir": "mert",
        // Extract content features using dataloader
        "pin_memory": true,
        "num_workers": 8,
        "content_feature_batch_size": 16,
        // Features used for model training
        "use_mel": true,
        "use_min_max_norm_mel": true,
        "use_frame_pitch": true,
        "use_uv": true,
        "use_frame_energy": true,
        "use_log_scale_pitch": false,
        "use_log_scale_energy": false,
        "use_spkid": true,
        // Meta file
        "train_file": "train.json",
        "valid_file": "test.json",
        "spk2id": "singers.json",
        "utt2spk": "utt2singer"
    },
    "model": {
        "teacher_model_path": "[Your Teacher Model Path].bin",
        "condition_encoder": {
            "merge_mode": "add",
            "input_melody_dim": 1,
            "use_log_f0": true,
            "n_bins_melody": 256,
            //# Quantization (0 for not quantization)
            "output_melody_dim": 384,
            "input_loudness_dim": 1,
            "use_log_loudness": true,
            "n_bins_loudness": 256,
            "output_loudness_dim": 384,
            "use_whisper": false,
            "use_contentvec": false,
            "use_wenet": false,
            "use_mert": false,
            "whisper_dim": 1024,
            "contentvec_dim": 256,
            "mert_dim": 256,
            "wenet_dim": 512,
            "content_encoder_dim": 384,
            "output_singer_dim": 384,
            "singer_table_size": 512,
            "output_content_dim": 384,
            "use_spkid": true
        },
        "comosvc": {
            "distill": false,
            // conformer encoder
            "input_dim": 384,
            "output_dim": 100,
            "n_heads": 2,
            "n_layers": 6,
            "filter_channels": 512,
            "dropout": 0.1,
            // karras diffusion
            "P_mean": -1.2,
            "P_std": 1.2,
            "sigma_data": 0.5,
            "sigma_min": 0.002,
            "sigma_max": 80,
            "rho": 7,
            "n_timesteps": 18,
        },
        "diffusion": {
            // Diffusion steps encoder
            "step_encoder": {
                "dim_raw_embedding": 128,
                "dim_hidden_layer": 512,
                "activation": "SiLU",
                "num_layer": 2,
                "max_period": 10000
            },
            // Diffusion decoder
            "model_type": "bidilconv",
            // bidilconv, unet2d, TODO: unet1d
            "bidilconv": {
                "base_channel": 384,
                "n_res_block": 20,
                "conv_kernel_size": 3,
                "dilation_cycle_length": 4,
                // specially, 1 means no dilation
                "conditioner_size": 100
            }
        },
    },
    "train": {
        // Basic settings
        "fast_steps": 0,
        "batch_size": 64,
        "gradient_accumulation_step": 1,
        "max_epoch": -1,
        // -1 means no limit
        "save_checkpoint_stride": [
            10,
            100
        ],
        // unit is epoch
        "keep_last": [
            3,
            -1
        ],
        // -1 means infinite, if one number will broadcast
        "run_eval": [
            false,
            true
        ],
        // if one number will broadcast
        // Fix the random seed
        "random_seed": 10086,
        // Batchsampler
        "sampler": {
            "holistic_shuffle": true,
            "drop_last": true
        },
        // Dataloader
        "dataloader": {
            "num_worker": 32,
            "pin_memory": true
        },
        // Trackers
        "tracker": [
            "tensorboard"
            // "wandb",
            // "cometml",
            // "mlflow",
        ],
        // Optimizer
        "optimizer": "AdamW",
        "adamw": {
            "lr": 5.0e-5
            // nn model lr
        },
        // LR Scheduler
        "scheduler": "ReduceLROnPlateau",
        "reducelronplateau": {
            "factor": 0.8,
            "patience": 10,
            // unit is epoch
            "min_lr": 5.0e-6
        }
    },
    "inference": {
        "comosvc": {
            "inference_steps": 40
        }
    }
}