{
  "schema_version": 1,
  "family": "breeze_tts",
  "display_name": "BreezeTTS 2",
  "description": "BreezeTTS 2 native GGUF package for instruction-conditioned text-to-speech and prompt-audio voice cloning with T5Gemma text conditioning, Qwen-style acoustic code generation, depth codebook decoding, and Mimi waveform decoding.",
  "category": "tts",
  "status": "supported",
  "tasks": [
    "tts",
    "clone",
    "design"
  ],
  "modes": [
    "offline",
    "streaming"
  ],
  "languages": [
    "zh",
    "en"
  ],
  "runtime": {
    "tags": [
      "gguf"
    ]
  },
  "dependencies": [],
  "capabilities": {
    "design": [
      "voice_design"
    ],
    "tts": [
      "style_control"
    ],
    "clone": [
      "speaker_reference"
    ]
  },
  "options": {
    "request": [
      {
        "name": "instruction",
        "type": "string",
        "description": "BreezeTTS generation instruction.",
        "required": false,
        "default": "Speak clearly and naturally."
      },
      {
        "name": "reference_text",
        "type": "string",
        "description": "Transcript for the prompt audio when cloning.",
        "required": false
      },
      {
        "name": "text_chunk_size",
        "type": "int",
        "description": "Maximum Unicode codepoints per long-form text chunk; default 600.",
        "required": false,
        "min": 1,
        "default": 600
      },
      {
        "name": "text_chunk_mode",
        "type": "enum",
        "description": "Framework text chunking mode.",
        "values": [
          "default",
          "tag_aware",
          "japanese",
          "endline"
        ],
        "required": false,
        "default": "default"
      },
      {
        "name": "max_tokens",
        "type": "int",
        "description": "Maximum generated BreezeTTS acoustic frames.",
        "required": false,
        "min": 1,
        "default": 1500
      },
      {
        "name": "guidance_scale",
        "type": "float",
        "description": "Classifier-free guidance scale.",
        "required": false,
        "min": 0.0,
        "default": 1.0
      },
      {
        "name": "temperature",
        "type": "float",
        "description": "Backbone first-codebook sampling temperature.",
        "required": false,
        "min": 0.0,
        "default": 0.9
      },
      {
        "name": "depth_temperature",
        "type": "float",
        "description": "Depth decoder codebook sampling temperature.",
        "required": false,
        "min": 0.0,
        "default": 0.9
      },
      {
        "name": "top_k",
        "type": "int",
        "description": "Top-k sampling limit; 0 disables top-k filtering.",
        "required": false,
        "min": 0,
        "default": 50
      },
      {
        "name": "top_p",
        "type": "float",
        "description": "Top-p sampling limit.",
        "required": false,
        "min": 0.0,
        "max": 1.0,
        "default": 1.0
      },
      {
        "name": "seed",
        "type": "int",
        "description": "Generation seed.",
        "required": false,
        "min": 0,
        "default": 0
      },
      {
        "name": "stream_frames_per_event",
        "type": "int",
        "description": "Generated codec frames per streaming audio event.",
        "required": false,
        "min": 1,
        "default": 16
      },
      {
        "name": "stream_lookahead_margin",
        "type": "int",
        "description": "Trailing codec frames held before emission to reduce streaming boundary artifacts.",
        "required": false,
        "min": 0,
        "default": 12
      }
    ],
    "session": [
      {
        "name": "weight_type",
        "type": "enum",
        "description": "BreezeTTS matmul weight storage type; default native.",
        "preset": "weight_type_full",
        "required": false,
        "default": "native"
      },
      {
        "name": "graph_arena_mb",
        "type": "int",
        "description": "Reusable runtime graph arena size in MiB; default 1024.",
        "required": false,
        "min": 1,
        "default": 1024
      },
      {
        "name": "weight_context_mb",
        "type": "int",
        "description": "Weight loading context size in MiB; default 2048.",
        "required": false,
        "min": 1,
        "default": 2048
      },
      {
        "name": "reference_cache_slots",
        "type": "int",
        "description": "Prepared reference-audio cache slots; set 0 to disable reuse.",
        "required": false,
        "min": 0,
        "default": 1
      },
      {
        "name": "attention",
        "type": "enum",
        "description": "Attention lowering; auto probes the backend and falls back to eager on GPUs without a flash kernel (e.g. sm70); default auto.",
        "required": false,
        "values": ["auto", "flash", "eager"],
        "default": "auto"
      },
      {
        "name": "bf16_activations",
        "type": "enum",
        "description": "Reference bf16 activation rounding as the official BreezeTTS 2 inference does it. auto = on for CUDA/HIP/Vulkan and off on Metal, where the cast cost is visible even with the fused round-to-bf16 op (~20% slower AR loop); on/off force it either way. On Metal, on also switches the KV cache to bf16 to match the reference implementation.",
        "required": false,
        "values": ["auto", "on", "off"],
        "default": "auto"
      }
    ],
    "load": []
  },
  "packages": [
    {
      "id": "breeze_tts_2_q4_0",
      "display_name": "BreezeTTS 2 Q4_0 GGUF",
      "default": false,
      "format": "gguf",
      "precision": "q4_0",
      "target_directory": "Breeze-TTS-2-GGUF",
      "files": [
        "Breeze-TTS-2-GGUF/breeze-tts-2-q4_0.gguf"
      ],
      "strip_prefix": "Breeze-TTS-2-GGUF",
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "audio-cpp/audio.cpp-gguf",
        "revision": "main",
        "gated": false
      }
    },
    {
      "id": "breeze_tts_2_q8_0",
      "display_name": "BreezeTTS 2 Q8_0 GGUF",
      "default": true,
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "Breeze-TTS-2-GGUF",
      "files": [
        "Breeze-TTS-2-GGUF/breeze-tts-2-q8_0.gguf"
      ],
      "strip_prefix": "Breeze-TTS-2-GGUF",
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "audio-cpp/audio.cpp-gguf",
        "revision": "main",
        "gated": false
      }
    },
    {
      "id": "breeze_tts_2_bf16",
      "display_name": "BreezeTTS 2 BF16 GGUF",
      "default": false,
      "format": "gguf",
      "precision": "bf16",
      "target_directory": "Breeze-TTS-2-GGUF",
      "files": [
        "Breeze-TTS-2-GGUF/breeze-tts-2-bf16.gguf"
      ],
      "strip_prefix": "Breeze-TTS-2-GGUF",
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "audio-cpp/audio.cpp-gguf",
        "revision": "main",
        "gated": false
      }
    }
  ],
  "ui": {
    "recommended_package": "breeze_tts_2_q8_0",
    "tags": [
      "TTS",
      "Clone",
      "Design",
      "GGUF"
    ],
    "docs": [
      "docs/tts.md",
      "docs/gguf.md"
    ]
  },
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "config_json": "model:config.json",
        "audio_tokenizer_config_json": "model:audio_tokenizer/config.json",
        "tokenizer_config_json": "model:tokenizer_config.json",
        "tokenizer_json": "model:tokenizer.json"
      },
      "tensors": {
        "model_weights": {
          "source": "weights:",
          "prefix": "model"
        }
      }
    }
  ]
}