{
  "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": "instruct",
        "type": "string",
        "description": "BreezeTTS generation instruction.",
        "required": false,
        "default": "Speak clearly and naturally."
      },
      {
        "name": "reference_text",
        "type": "string",
        "description": "Transcript for prompt/reference audio; required when reference audio codes are used, for either tts or clon.",
        "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. Must be > 0.",
        "required": false,
        "min": 0.0,
        "default": 0.9
      },
      {
        "name": "depth_temperature",
        "type": "float",
        "description": "Depth decoder codebook sampling temperature. Must be > 0.",
        "required": false,
        "min": 0.0,
        "default": 0.9
      },
      {
        "name": "top_k",
        "type": "int",
        "description": "Top-k sampling limit; 0 disables top-k filtering. Nonpositive values disable top-k filtering.",
        "required": false,
        "default": 50
      },
      {
        "name": "top_p",
        "type": "float",
        "description": "Positive nucleus sampling cutoff; values >=1 disable nucleus filtering.",
        "required": false,
        "min": 0,
        "default": 1
      },
      {
        "name": "seed",
        "type": "int",
        "description": "Generation seed.",
        "required": false,
        "min": 0,
        "default": 0
      },
      {
        "name": "stream_frames_per_event",
        "type": "int",
        "description": "Maximum generated codec frames per streaming audio event; fixed size with stream_chunk_policy=fixed.",
        "required": false,
        "min": 1,
        "default": 16
      },
      {
        "name": "stream_lookahead_margin",
        "type": "int",
        "description": "Additional trailing codec frames buffered before streaming emission. The decoder is causal; zero minimizes latency.",
        "required": false,
        "min": 0,
        "default": 0
      },
      {
        "name": "stream_chunk_policy",
        "type": "enum",
        "description": "Streaming chunk schedule: grow emits 1, 2, 4, ... frames up to stream_frames_per_event; fixed uses that size from the first event.",
        "values": ["grow", "fixed"],
        "required": false,
        "default": "grow"
      }
    ],
    "session": [
      {
        "name": "weight_type",
        "type": "enum",
        "description": "BreezeTTS matmul weight storage type; default native.",
        "required": false,
        "default": "native",
        "values": [
          "native",
          "f32",
          "f16",
          "bf16",
          "q8_0",
          "q4_0",
          "q4_k"
        ]
      },
      {
        "name": "graph_arena_mb",
        "type": "int",
        "description": "Reusable runtime graph arena size in MiB; default 32.",
        "required": false,
        "min": 1,
        "default": 32
      },
      {
        "name": "weight_context_mb",
        "type": "int",
        "description": "Weight loading context size in MiB; default 4.",
        "required": false,
        "min": 1,
        "default": 4
      },
      {
        "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"
      },
      {
        "name": "perf_mode",
        "type": "enum",
        "description": "AR decode preset. off uses the normal defaults (fused_depth, direct_prefill and skip_cache_round enabled on CUDA, Vulkan and Metal); cuda_perf additionally enables batched_cfg on CUDA. Explicit session options override the preset.",
        "required": false,
        "values": ["off", "cuda_perf"],
        "default": "off"
      },
      {
        "name": "fused_depth",
        "type": "bool",
        "description": "Decode each frame's depth codebooks in one backend graph with device-side CFG mixing and sampling. Device sampling uses the per-codebook path's torch random draws, so CUDA output normally matches it token for token. Vulkan/Metal sampling uses a different draw mapping. CUDA, Vulkan and Metal only.",
        "required": false,
        "default": true
      },
      {
        "name": "direct_prefill",
        "type": "bool",
        "description": "Offline generation prefills the backbone straight into the decode cache, as streaming does, instead of exporting the KV state to the host and importing it back. Same computation. CUDA, Vulkan and Metal only.",
        "required": false,
        "default": true
      },
      {
        "name": "skip_cache_round",
        "type": "bool",
        "description": "Skip re-rounding the F16 KV cache to bf16 after each update, which is an identity because the K/V rows are already bf16-rounded. Applies only with the bf16 reference path and flash attention. CUDA, Vulkan and Metal only.",
        "required": false,
        "default": true
      },
      {
        "name": "round_bf16_fusion",
        "type": "bool",
        "description": "Fold the bf16 activation rounding into the CUDA RMS-norm, add and SwiGLU kernels that produce it. Bit-identical to the unfused graph, with fewer kernel launches. CUDA only.",
        "required": false,
        "default": false
      },
      {
        "name": "batched_cfg",
        "type": "bool",
        "description": "With guidance_scale != 1, decode the conditional and unconditional backbone branches as one batch-2 step. Batching changes matmul accumulation order, so output differs from the two-runtime path. CUDA only.",
        "required": false,
        "default": false
      }
    ],
    "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"
        }
      }
    }
  ]
}