{
  "schema_version": 1,
  "family": "soprano_tts",
  "display_name": "Soprano",
  "description": "Soprano is an ultra-lightweight (~80M) English-only text-to-speech model. Syntax uses a 17-layer Qwen3-style causal LM (hidden 512, vocab 8192) that autoregressively emits per-frame 512-dimensional features; a non-iterative Vocos-style decoder (ConvNeXt backbone + single ISTFT head, n_fft 2048 / hop 512) turns those features into 32 kHz audio. No diffusion refinement is performed in the decoder.",
  "category": "tts",
  "status": "community",
  "tasks": [
    "tts"
  ],
  "modes": [
    "offline",
    "streaming"
  ],
  "languages": [
    "en"
  ],
  "runtime": {
    "tags": [
      "gguf",
      "stream"
    ]
  },
  "capabilities": {
    "tts": [
      "long_form"
    ]
  },
  "options": {
    "request": [
      {
        "name": "max_tokens",
        "type": "int",
        "description": "Maximum generated audio frames for the autoregressive LM; default 512.",
        "required": false,
        "min": 1,
        "default": 512
      },
      {
        "name": "temperature",
        "type": "float",
        "description": "Autoregressive sampling temperature; must be > 0, default 0.3.",
        "required": false,
        "min": 0.0,
        "default": 0.3
      },
      {
        "name": "top_p",
        "type": "float",
        "description": "Nucleus sampling probability; default 0.95. Values >=1 disable nucleus filtering.",
        "required": false,
        "min": 0,
        "default": 0.95
      },
      {
        "name": "repetition_penalty",
        "type": "float",
        "description": "Repetition penalty applied to the LM head; default 1.2. Must be positive; values below 1 are accepted.",
        "required": false,
        "default": 1.2
      },
      {
        "name": "eos_bias",
        "type": "float",
        "description": "Additive bias on the EOS token logit during generation. Positive values make the model stop sooner when speech ends (mitigating runaway generations that hit max_tokens); negative values encourage longer utterances. Default 0 disables the adjustment.",
        "required": false,
        "default": 0.0
      },
      {
        "name": "seed",
        "type": "int",
        "description": "Autoregressive sampling seed; omitted requests choose a random seed.",
        "required": false,
        "min": 0
      }
    ],
    "session": [
      {
        "name": "text_chunk_size",
        "type": "int",
        "description": "Maximum codepoints per sentence chunk before the model generates and decodes separately. Smaller values keep prompts short (more reliable EOS) but increase overhead. Default 200.",
        "required": false,
        "min": 1,
        "default": 200
      },
      {
        "name": "backbone_weight_type",
        "type": "enum",
        "preset": "weight_type_full",
        "required": false,
        "default": "f32",
        "description": "Storage type for the Qwen3 LM backbone weights."
      },
      {
        "name": "decoder_weight_type",
        "type": "enum",
        "preset": "weight_type_conv",
        "required": false,
        "default": "f32",
        "description": "Storage type for the Vocos decoder weights."
      }
    ],
    "load": []
  },
  "package_defaults": {
    "download": {
      "kind": "huggingface_snapshot",
      "repo": "WalkingCat/Soprano-1.1-80M-GGUF",
      "revision": "main",
      "gated": false
    }
  },
  "packages": [
    {
      "id": "soprano_1_1_80m_q8_0",
      "display_name": "Soprano-1.1-80M Q8_0 GGUF",
      "default": true,
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "Soprano-1.1-80M-GGUF",
      "files": [
        "Soprano-1.1-80M-GGUF/soprano-1.1-80m-q8_0.gguf"
      ],
      "strip_prefix": "Soprano-1.1-80M-GGUF"
    },
    {
      "id": "soprano_1_1_80m_bf16",
      "display_name": "Soprano-1.1-80M BF16 GGUF",
      "format": "gguf",
      "precision": "bf16",
      "target_directory": "Soprano-1.1-80M-GGUF",
      "files": [
        "Soprano-1.1-80M-GGUF/soprano-1.1-80m-bf16.gguf"
      ],
      "strip_prefix": "Soprano-1.1-80M-GGUF"
    }
  ],
  "dependencies": [],
  "ui": {
    "recommended_package": "soprano_1_1_80m_q8_0",
    "tags": [
      "TTS",
      "Stream"
    ],
    "docs": [
      "docs/models/soprano_tts.md"
    ]
  },
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "config": "model:config.json",
        "generation_config": "model:generation_config.json",
        "tokenizer_json": "model:tokenizer.json"
      },
      "tensors": {
        "backbone": "weights:",
        "decoder": "weights:"
      }
    },
    {
      "format": "safetensors",
      "roots": {
        "model": "."
      },
      "files": {
        "config": "model:config.json",
        "generation_config": "model:generation_config.json",
        "tokenizer_json": "model:tokenizer.json"
      },
      "tensors": {
        "backbone": "model:combined.safetensors",
        "decoder": "model:combined.safetensors"
      }
    }
  ]
}