{
  "schema_version": 1,
  "family": "zipvoice",
  "display_name": "ZipVoice",
  "description": "Community ZipVoice / ZipVoice-Distill flow-matching TTS with a TTSZipformer backbone (U-Net downsampling stacks, compact relative position attention), duration prediction from prompt ratio, Euler solver with t-shift and classifier-free guidance, and Vocos mel-24kHz vocoder (k2-fsa/ZipVoice).",
  "category": "tts",
  "status": "community",
  "tasks": [
    "tts",
    "clone"
  ],
  "modes": [
    "offline",
    "streaming"
  ],
  "languages": [
    "en",
    "zh"
  ],
  "runtime": {
    "tags": [
      "server",
      "stream"
    ]
  },
  "capabilities": {
    "clone": [
      "speaker_reference"
    ]
  },
  "options": {
    "request": [
      {
        "name": "reference_text",
        "type": "string",
        "description": "Transcript matching the reference voice audio; required for zero-shot cloning.",
        "required": true
      },
      {
        "name": "text_chunk_size",
        "type": "int",
        "description": "Maximum characters per generated text chunk; default 128.",
        "required": false,
        "min": 1,
        "default": 128
      },
      {
        "name": "text_chunk_mode",
        "type": "enum",
        "description": "Framework text chunking mode.",
        "preset": "text_chunk_mode_full",
        "required": false,
        "default": "word_budget"
      },
      {
        "name": "guidance_scale",
        "type": "float",
        "description": "Guidance scale; inherits the session default (3.0 for both variants). ZipVoice-Distill uses a guidance-scale embedding; ZipVoice uses batched CFG. Zero disables guidance. Any finite nonnegative value is accepted.",
        "required": false,
        "min": 0,
        "default": 3
      },
      {
        "name": "num_inference_steps",
        "type": "int",
        "description": "Euler ODE steps; inherits the session default (8 for both variants).",
        "required": false,
        "min": 1,
        "max": 64,
        "default": 8
      },
      {
        "name": "shift",
        "type": "float",
        "description": "Shift timesteps toward low SNR (smaller = stronger shift); default 0.5. Must be finite and positive.",
        "required": false,
        "default": 0.5
      },
      {
        "name": "speed",
        "type": "float",
        "description": "Speech speed multiplier applied through the prompt-duration ratio; default 1.0. Must be finite and positive.",
        "required": false,
        "default": 1
      },
      {
        "name": "feat_scale",
        "type": "float",
        "description": "Feature scale applied to log-mel features; default 0.1 (reference default). Must be finite and positive.",
        "required": false,
        "default": 0.1
      },
      {
        "name": "target_rms",
        "type": "float",
        "description": "Target RMS for prompt loudness normalization; 0 disables. Default 0.1. Negative values disable normalization; values above 1 are accepted.",
        "required": false,
        "default": 0.1
      },
      {
        "name": "seed",
        "type": "int",
        "description": "Noise seed; default 666 (reference default).",
        "required": false,
        "min": 0,
        "default": 666
      },
      {
        "name": "language",
        "type": "string",
        "description": "espeak language for phonemization; default en-us.",
        "required": false,
        "default": "en-us"
      }
    ],
    "session": [
      {
        "name": "vocos_path",
        "type": "string",
        "description": "Path to the Vocos vocoder checkpoint (vocos.safetensors or GGUF); required unless bundled in the model GGUF or placed next to the checkpoint.",
        "required": false
      },
      {
        "name": "guidance_scale",
        "type": "float",
        "description": "Default guidance scale for requests that do not set it; must be finite and nonnegative.",
        "required": false,
        "min": 0.0,
        "default": 3.0
      },
      {
        "name": "num_inference_steps",
        "type": "int",
        "description": "Default Euler steps for requests that do not set it.",
        "required": false,
        "min": 1,
        "max": 64,
        "default": 8
      },
      {
        "name": "shift",
        "type": "float",
        "description": "Default timestep shift for requests that do not set it; must be finite and positive.",
        "required": false,
        "default": 0.5
      },
      {
        "name": "espeak_library_path",
        "type": "string",
        "description": "Path to the espeak-ng shared library (e.g. /opt/homebrew/lib/libespeak-ng.dylib) for tokenizer=espeak.",
        "required": false
      },
      {
        "name": "espeak_data_path",
        "type": "string",
        "description": "Path to the espeak-ng data directory or package for tokenizer=espeak.",
        "required": false
      }
    ],
    "load": []
  },
  "package_defaults": {
    "download": {
      "kind": "huggingface_snapshot",
      "repo": "davidxifeng/zipvoice-gguf",
      "revision": "main",
      "gated": false
    }
  },
  "packages": [
    {
      "id": "zipvoice_distill_gguf",
      "display_name": "ZipVoice-Distill GGUF (local conversion)",
      "description": "Self-contained GGUF: flow-matching model, bundled Vocos vocoder and the embedded text-frontend sidecars (tokens, config, zh tables) in one file. Converted from k2-fsa/ZipVoice with export_zipvoice_zh_dict.py + convert_zipvoice.py.",
      "default": true,
      "format": "gguf",
      "precision": "orig",
      "target_directory": "ZipVoice-Distill-GGUF",
      "files": [
        "zipvoice-distill-orig.gguf"
      ]
    },
    {
      "id": "zipvoice_distill_q8_0",
      "display_name": "ZipVoice-Distill GGUF (Q8_0)",
      "description": "Self-contained Q8_0 GGUF with bundled Vocos and embedded text-frontend sidecars.",
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "ZipVoice-Distill-GGUF",
      "files": [
        "zipvoice-distill-q8_0.gguf"
      ]
    }
  ],
  "dependencies": [],
  "ui": {
    "recommended_package": "zipvoice_distill_gguf",
    "tags": [
      "TTS",
      "Clone"
    ],
    "docs": [
      "docs/models/zipvoice.md"
    ]
  },
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "tokens": "model:tokens.txt",
        "model_config": "model:model.json"
      },
      "optional_files": {
        "zh_chars": "model:zh_chars.tsv",
        "zh_phrases": "model:zh_phrases.tsv",
        "zh_syllables": "model:zh_syllables.tsv",
        "zh_jieba_dict": "model:zh_jieba_dict.txt",
        "zh_hmm_model": "model:zh_hmm_model.txt"
      },
      "tensors": {
        "model": {
          "source": "weights:",
          "prefix": "model"
        }
      },
      "optional_tensors": {
        "vocos_vocoder": {
          "source": "weights:",
          "prefix": "vocos"
        }
      }
    }
  ]
}