{
  "family": "vibevoice",
  "display_name": "VibeVoice",
  "description": "Microsoft long-form multi-speaker TTS model for expressive conversational audio such as podcasts, supporting up to 90 minutes of speech with as many as four speakers.",
  "category": "tts",
  "status": "supported",
  "tasks": [
    "tts"
  ],
  "modes": [
    "offline"
  ],
  "languages": [
    "en",
    "zh"
  ],
  "capabilities": {
    "tts": [
      "multi_speaker",
      "long_form"
    ]
  },
  "options": {
    "request": [
      {
        "name": "voice_samples",
        "type": "audio_path_list",
        "description": "Comma-separated speaker reference WAV paths for up to four multi-speaker prompts; reference audio is trimmed to 10 seconds on CPU/Vulkan/Metal and 30 seconds on CUDA.",
        "required": false
      },
      {
        "name": "max_tokens",
        "type": "int",
        "description": "Maximum decoder generation tokens; default 0 leaves generation capped by model context and max_length_times.",
        "required": false,
        "min": 0,
        "default": 0
      },
      {
        "name": "num_inference_steps",
        "type": "int",
        "description": "Diffusion inference steps; defaults to model config, usually 10.",
        "required": false,
        "min": 1
      },
      {
        "name": "guidance_scale",
        "type": "float",
        "description": "Classifier-free guidance scale; default 1.3.",
        "required": false,
        "min": 0.0,
        "default": 1.3
      },
      {
        "name": "max_length_times",
        "type": "float",
        "description": "Maximum decoder generation steps as a positive multiplier of prompt token steps; default 2.0.",
        "required": false,
        "default": 2.0
      },
      {
        "name": "do_sample",
        "type": "bool",
        "description": "Sample constrained speech-control tokens instead of greedy selection; default false.",
        "required": false,
        "default": false
      },
      {
        "name": "temperature",
        "type": "float",
        "description": "Constrained-token sampling temperature; must be positive when sampling, default 1.0.",
        "required": false,
        "default": 1.0
      },
      {
        "name": "top_k",
        "type": "int",
        "description": "Top-k filter for constrained-token sampling; 0 disables top-k filtering, default 50.",
        "required": false,
        "min": 0,
        "default": 50
      },
      {
        "name": "top_p",
        "type": "float",
        "description": "Nucleus filter for constrained-token sampling; must be in (0, 1], default 1.0.",
        "required": false,
        "min": 0.0,
        "max": 1.0,
        "default": 1.0
      },
      {
        "name": "seed",
        "type": "int",
        "description": "Random seed for prompt acoustic sampling, diffusion noise, and optional constrained-token sampling; default 1234.",
        "required": false,
        "min": 0,
        "default": 1234
      },
      {
        "name": "prompt_noise_path",
        "type": "path",
        "description": "F32 prompt acoustic noise file for deterministic reference-voice acoustic latent replay.",
        "required": false
      },
      {
        "name": "diffusion_noise_path",
        "type": "path",
        "description": "F32 initial diffusion noise file for deterministic generated-speech diffusion replay.",
        "required": false
      }
    ],
    "session": [
      {
        "name": "weight_type",
        "type": "enum",
        "description": "Tokenizer, connector, decoder, and diffusion-head weight storage type; default native.",
        "preset": "weight_type_full",
        "required": false,
        "default": "native"
      },
      {
        "name": "tokenizer_weight_type",
        "type": "enum",
        "description": "Audio tokenizer weight storage type; defaults to weight_type when set, otherwise native.",
        "preset": "weight_type_full",
        "required": false
      },
      {
        "name": "connector_weight_type",
        "type": "enum",
        "description": "Acoustic and semantic connector weight storage type; defaults to weight_type when set, otherwise native.",
        "preset": "weight_type_full",
        "required": false
      },
      {
        "name": "decoder_weight_type",
        "type": "enum",
        "description": "Language decoder weight storage type; defaults to weight_type when set, otherwise native.",
        "preset": "weight_type_full",
        "required": false
      },
      {
        "name": "diffusion_head_weight_type",
        "type": "enum",
        "description": "Diffusion prediction head weight storage type; defaults to weight_type when set, otherwise native.",
        "preset": "weight_type_full",
        "required": false
      }
    ],
    "load": [
      {
        "name": "lora_path",
        "type": "path",
        "description": "Fine-tune adapter dir (LM LoRA + diffusion head + acoustic/semantic connectors) merged at load time.",
        "required": false
      },
      {
        "name": "lora_scale",
        "type": "float",
        "description": "LoRA merge scale override; defaults to lora_alpha / r.",
        "required": false,
        "min": 0.0
      }
    ]
  },
  "runtime": {
    "tags": [
      "gguf"
    ]
  },
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "config": "model:config.json",
        "preprocessor_config": "model:preprocessor_config.json",
        "tokenizer_config": "model:tokenizer_config.json",
        "tokenizer_json": "model:tokenizer.json",
        "tokenizer_vocab": "model:vocab.json",
        "tokenizer_merges": "model:merges.txt"
      },
      "tensors": {
        "model_weights": "weights:"
      }
    },
    {
      "format": "safetensors",
      "roots": {
        "model": "."
      },
      "files": {
        "config": "model:config.json",
        "preprocessor_config": "model:preprocessor_config.json",
        "tokenizer_config": "model:tokenizer_config.json",
        "tokenizer_json": "model:tokenizer.json",
        "tokenizer_vocab": "model:vocab.json",
        "tokenizer_merges": "model:merges.txt"
      },
      "tensors": {
        "model_weights": "model:model.safetensors.index.json"
      }
    }
  ],
  "package_defaults": {
    "download": {
      "kind": "huggingface_snapshot",
      "repo": "audio-cpp/audio.cpp-gguf",
      "revision": "main",
      "gated": false
    }
  },
  "packages": [
    {
      "id": "vibevoice_1_5b_q8_0",
      "display_name": "VibeVoice 1.5B Q8_0 GGUF",
      "default": true,
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "VibeVoice-1.5B-GGUF",
      "files": [
        "VibeVoice-1.5B-GGUF/vibevoice-1.5b-q8_0.gguf"
      ],
      "strip_prefix": "VibeVoice-1.5B-GGUF"
    },
    {
      "id": "vibevoice_1_5b_bf16",
      "display_name": "VibeVoice 1.5B BF16 GGUF",
      "format": "gguf",
      "precision": "bf16",
      "target_directory": "VibeVoice-1.5B-GGUF",
      "files": [
        "VibeVoice-1.5B-GGUF/vibevoice-1.5b-bf16.gguf"
      ],
      "strip_prefix": "VibeVoice-1.5B-GGUF"
    }
  ]
}