configs/config_24k.json

{
    "semantic_llm": {
        "start_text_token": 32000,
        "stop_text_token": 32001,
        "num_text_tokens": 32002,
        "start_audio_token": 16384,
        "stop_audio_token": 16385,
        "num_audio_tokens": 16386,
        "llm_hidden_size": 1024,
        "llm_intermediate_size": 4096,
        "llm_num_layers": 30,
        "llm_num_heads": 16,
        "llm_max_audio_seq_len": 630,
        "llm_max_text_seq_len": 402,
        "llm_max_prompt_len": 250,
        "code_stride_len": 640,
        "EOS_TOKEN": 16385
    },
    "acoustic_llm": {
        "n_stacks": 1,
        "layers": 24,
        "model_dim": 1536,
        "heads": 16,
        "max_text_tokens": 2048,
        "max_speech_tokens": 2048,
        "max_conditioning_inputs": 1,
        "number_text_tokens": 16386,
        "start_text_token": 16384,
        "stop_text_token": 16385,
        "n_frames_per_step": 1,
        "n_heads_per_frame": 8,
        "delay_prediction": 1,
        "upsample_factors": 1,
        "streaming_delayed_frames": 8,
        "number_speech_tokens": 16386,
        "start_speech_token": 16384,
        "stop_speech_token": 16385,
        "speaker_embedding_pretrained": true,
        "speaker_embedding_ckpt": null,
        "speaker_embedding_dim": 512,
        "temperature": 0.5,
        "repetition_penalty": 2.0,
        "top_p": 0.5,
        "top_k": 25
    },
    "acoustic_codec": {
        "n_model_size": 1024,
        "encoder_config": {
            "ngf": 48,
            "up_ratios": [
                2,
                4,
                4,
                4,
                5
            ],
            "causal": true
        },
        "decoder_config": {
            "upsample_initial_channel": 1536,
            "ngf": 48,
            "up_ratios": [
                6,
                5,
                4,
                4,
                2
            ],
            "causal": true
        },
        "vq_config": {
            "n_groups": 8,
            "ordered": true,
            "codebook_size": [
                128,
                128,
                128,
                128,
                128,
                128,
                128,
                128,
                128,
                128,
                128,
                128,
                128,
                128,
                128,
                128
            ],
            "codebook_dim": [
                8,
                8,
                8,
                8,
                8,
                8,
                8,
                8,
                8,
                8,
                8,
                8,
                8,
                8,
                8,
                8
            ],
            "requires_projection": true,
            "decay": 0.99,
            "threshold_ema_dead_code": 0,
            "commitment_weight": 0.01
        },
        "resampler_config": {
            "source_sr": 16000,
            "target_sr": 16000
        }
    },
    "semantic_tokenizer": {
        "in_dim": 1024,
        "out_dim": 80,
        "n_model_size": 512,
        "downsample_scales": [
            1,
            1,
            1,
            2
        ],
        "upsample_scales": [
            [
                2,
                1
            ],
            [
                2,
                1,
                1,
                1
            ]
        ],
        "mel_config": {
            "style": "BigVGAN",
            "filter_length": 1024,
            "hop_length": 160,
            "win_length": 640,
            "n_mel_channels": 80,
            "sampling_rate": 16000
        },
        "vq_config": {
            "codebook_size": [
                128,
                128
            ],
            "codebook_dim": [
                128,
                128
            ],
            "requires_projection": true
        },
        "tree_config": [
            {
                "downsample_rate": 1,
                "n_groups": 1,
                "dropout": 0
            }
        ],
        "n_samples_per_token": 640,
        "checkpointing": true
    }
}