voice_audio_multi/model_specs/sopro_tts.json

{
  "schema_version": 1,
  "family": "sopro_tts",
  "display_name": "Sopro V2 Turbo",
  "description": "Community Sopro V2 Turbo (samuel-vitorino/sopro-v2-turbo): a 120M zero-shot voice-cloning TTS. SentencePiece text tokenizer, style-prefix conditioned autoregressive semantic LM over FSQ tokens, rectified-flow acoustic DiT and a Vocos ISTFT vocoder at 24 kHz.",
  "category": "tts",
  "status": "community",
  "tasks": [
    "tts",
    "clone"
  ],
  "modes": [
    "offline",
    "streaming"
  ],
  "languages": [
    "en",
    "pt",
    "fr",
    "de"
  ],
  "runtime": {
    "tags": [
      "server"
    ]
  },
  "capabilities": {
    "clone": [
      "speaker_reference"
    ]
  },
  "options": {
    "request": [
      {
        "name": "language",
        "type": "string",
        "description": "Language tag (en, pt, fr, de); omitted inherits transcript language, then the session default.",
        "required": false
      },
      {
        "name": "temperature",
        "type": "float",
        "description": "Semantic sampling temperature; nonpositive selects greedy decoding. Omitted uses checkpoint generation.temperature.",
        "required": false
      },
      {
        "name": "top_p",
        "type": "float",
        "description": "Semantic nucleus cutoff; values outside [0,1] are disabled/clamped by the sampler. Omitted uses checkpoint generation.top_p.",
        "required": false
      },
      {
        "name": "top_k",
        "type": "int",
        "description": "Semantic top-k; nonpositive disables filtering. Omitted uses checkpoint generation.top_k.",
        "required": false
      },
      {
        "name": "num_inference_steps",
        "type": "int",
        "description": "Positive acoustic sampling steps; omitted uses checkpoint generation.steps.",
        "required": false,
        "min": 1
      },
      {
        "name": "max_duration_sec",
        "type": "float",
        "description": "Positive maximum segment duration; must be >= min_duration_sec. Omitted uses checkpoint generation.max_seconds.",
        "required": false
      },
      {
        "name": "min_duration_sec",
        "type": "float",
        "description": "Nonnegative minimum duration; must be <= max_duration_sec. Omitted uses checkpoint generation.min_seconds.",
        "required": false,
        "min": 0
      },
      {
        "name": "reference_duration_sec",
        "type": "float",
        "description": "Positive reference window in seconds; omitted uses checkpoint generation.ref_seconds.",
        "required": false
      },
      {
        "name": "text_chunk_size",
        "type": "int",
        "description": "Positive codepoint budget per segment; omitted uses checkpoint generation.max_segment_chars.",
        "required": false,
        "min": 1
      },
      {
        "name": "seed",
        "type": "int",
        "description": "Non-negative seed for semantic sampling and the acoustic noise prior; omit for a random seed.",
        "required": false,
        "min": 0
      }
    ],
    "session": [
      {
        "name": "language",
        "type": "string",
        "description": "Default language tag for requests that do not set one.",
        "required": false,
        "default": ""
      },
      {
        "name": "voice_cache_slots",
        "type": "int",
        "description": "Reference voices kept prepared between requests, so a repeated voice skips the encoders and, when streaming, the prompt solve. Output is identical with or without it. A streamed 10 s voice holds about 60 MB. Default 0 (off).",
        "required": false,
        "min": 0,
        "default": 0
      },
      {
        "name": "matmul_weight_type",
        "type": "enum",
        "preset": "weight_type_full",
        "description": "Storage type for the matmul weights of every stage (semantic LM, acoustic DiT, encoders, vocoder head).",
        "required": false,
        "default": "f32"
      },
      {
        "name": "conv_weight_type",
        "type": "enum",
        "preset": "weight_type_conv",
        "description": "Storage type for convolution weights (speaker/semantic encoders and the Vocos backbone).",
        "required": false,
        "default": "f32"
      }
    ],
    "load": []
  },
  "packages": [
    {
      "id": "sopro_v2_turbo_f16",
      "display_name": "Sopro V2 Turbo F16 GGUF",
      "description": "Single-file GGUF of all four stages with the config and tokenizer embedded: f16 weights, mel front ends kept in f32.",
      "default": true,
      "format": "gguf",
      "precision": "f16",
      "target_directory": "sopro-v2-turbo-GGUF",
      "files": [
        "sopro-v2-turbo-f16.gguf"
      ],
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "samuel-vitorino/sopro-v2-turbo",
        "revision": "main",
        "gated": false
      }
    },
    {
      "id": "sopro_v2_turbo_safetensors",
      "display_name": "Sopro V2 Turbo (upstream safetensors)",
      "description": "Upstream checkpoint from samuel-vitorino/sopro-v2-turbo: config.json, tokenizer.model and the four safetensors stages. Runs directly, no conversion needed.",
      "format": "safetensors",
      "precision": "orig",
      "target_directory": "sopro-v2-turbo",
      "files": [
        "config.json",
        "tokenizer.model",
        "model.safetensors",
        "semantic_encoder.safetensors",
        "speaker_encoder.safetensors",
        "vocoder.safetensors"
      ],
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "samuel-vitorino/sopro-v2-turbo",
        "revision": "main",
        "gated": false
      }
    }
  ],
  "dependencies": [],
  "ui": {
    "recommended_package": "sopro_v2_turbo_safetensors",
    "tags": [
      "TTS",
      "Clone"
    ],
    "docs": [
      "docs/models/sopro_tts.md"
    ]
  },
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "config": "model:config.json",
        "tokenizer": "model:tokenizer.model"
      },
      "tensors": {
        "model": {
          "source": "weights:",
          "prefix": "model"
        },
        "semantic_encoder": {
          "source": "weights:",
          "prefix": "semantic_encoder"
        },
        "speaker_encoder": {
          "source": "weights:",
          "prefix": "speaker_encoder"
        },
        "vocoder": {
          "source": "weights:",
          "prefix": "vocoder"
        }
      }
    },
    {
      "format": "safetensors",
      "roots": {
        "model": "."
      },
      "files": {
        "config": "model:config.json",
        "tokenizer": "model:tokenizer.model"
      },
      "tensors": {
        "model": {
          "source": "model:model.safetensors"
        },
        "semantic_encoder": {
          "source": "model:semantic_encoder.safetensors"
        },
        "speaker_encoder": {
          "source": "model:speaker_encoder.safetensors"
        },
        "vocoder": {
          "source": "model:vocoder.safetensors"
        }
      }
    }
  ]
}