{
  "schema_version": 1,
  "family": "vibeasr",
  "display_name": "VibeVoice-ASR-BitNet",
  "description": "VibeASR.cpp's CPU-first VibeVoice ASR port: an INT8 (I8_S) audio VAE encoder feeding a ternary (I2_S) Qwen2 decoder, ported to audio.cpp.",
  "category": "asr",
  "status": "community",
  "tasks": [
    "asr"
  ],
  "modes": [
    "offline"
  ],
  "languages": [
    "en",
    "zh",
    "fr",
    "it",
    "ko",
    "pt",
    "vi"
  ],
  "capabilities": {},
  "options": {
    "request": [
      {
        "name": "output_format",
        "type": "enum",
        "description": "Prompt suffix asked of the decoder: plain transcription text, or JSON rows with Start/End/Speaker/Content.",
        "values": [
          "text",
          "json"
        ],
        "required": false,
        "default": "text"
      },
      {
        "name": "context",
        "type": "string",
        "description": "Extra context injected into the prompt (names, jargon) to bias the transcription.",
        "required": false,
        "default": ""
      },
      {
        "name": "max_tokens",
        "type": "int",
        "description": "Cap on decoded tokens for one request.",
        "required": false,
        "min": 1,
        "default": 1024
      }
    ],
    "session": [
      {
        "name": "encoder_graph_arena_mb",
        "type": "int",
        "description": "VAE encoder graph arena size in MB. Must be positive; sufficient capacity depends on the graph.",
        "required": false,
        "min": 1,
        "default": 64
      },
      {
        "name": "prefill_graph_arena_mb",
        "type": "int",
        "description": "Decoder prefill graph arena size in MB. Must be positive; sufficient capacity depends on the graph.",
        "required": false,
        "min": 1,
        "default": 256
      },
      {
        "name": "decode_graph_arena_mb",
        "type": "int",
        "description": "Decoder single-step graph arena size in MB. Must be positive; sufficient capacity depends on the graph.",
        "required": false,
        "min": 1,
        "default": 256
      }
    ],
    "load": []
  },
  "runtime": {
    "tags": [
      "gguf",
      "cpu"
    ]
  },
  "packages": [
    {
      "id": "vibeasr_bitnet_i2_s",
      "display_name": "VibeVoice-ASR-BitNet I8_S encoder + I2_S decoder",
      "description": "Upstream VibeASR.cpp GGUF package. The two GGUFs carry the VibeASR ggml fork's type ids and need one pass of tools/community_models/convert_vibeasr_gguf.py --in-place before audio.cpp can load them.",
      "default": true,
      "format": "gguf",
      "precision": "native",
      "target_directory": "VibeVoice-ASR-BitNet",
      "files": [
        "vibeasr-vae-encoder-i8_s.gguf",
        "vibeasr-lm-i2_s-embed-q6_k.gguf",
        "tokenizer.json",
        "tokenizer_config.json"
      ],
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "microsoft/VibeVoice-ASR-BitNet",
        "revision": "main",
        "gated": false
      }
    }
  ],
  "dependencies": [],
  "ui": {
    "recommended_package": "vibeasr_bitnet_i2_s",
    "tags": [
      "ASR",
      "GGUF"
    ],
    "docs": [
      "docs/models/vibeasr.md"
    ],
    "summary": "INT8 encoder plus ternary Qwen2 decoder transcription on CPU."
  },
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": "."
      },
      "files": {
        "tokenizer_json": "model:tokenizer.json",
        "tokenizer_config": "model:tokenizer_config.json"
      },
      "optional_files": {},
      "tensors": {
        "vae_weights": "model:vibeasr-vae-encoder-i8_s.gguf",
        "lm_weights": "model:vibeasr-lm-i2_s-embed-q6_k.gguf"
      }
    }
  ]
}