voice_audio_multi/model_specs/mira_tts.json

{
  "schema_version": 1,
  "family": "mira_tts",
  "display_name": "MiraTTS",
  "description": "MiraTTS is a community voice-cloning TTS model with a Qwen2 autoregressive speech-token generator, ECAPA/Perceiver speaker tokenizer, conditional acoustic processor, and DAC decoder.",
  "category": "tts",
  "status": "experimental",
  "tasks": [
    "tts",
    "clone"
  ],
  "modes": [
    "offline",
    "streaming"
  ],
  "languages": [
    "en"
  ],
  "runtime": {
    "tags": [
      "gguf"
    ]
  },
  "capabilities": {
    "tts": [
      "speaker_reference"
    ],
    "clone": [
      "speaker_reference"
    ]
  },
  "options": {
    "request": [
      {
        "name": "max_tokens",
        "type": "int",
        "description": "Maximum autoregressive speech tokens; default 1024.",
        "required": false,
        "min": 1,
        "default": 1024
      },
      {
        "name": "temperature",
        "type": "float",
        "description": "Sampling temperature; default 0.8. Must be positive.",
        "required": false,
        "default": 0.8
      },
      {
        "name": "top_k",
        "type": "int",
        "description": "Top-k sampling cutoff; default 50.",
        "required": false,
        "min": 1,
        "default": 50
      },
      {
        "name": "top_p",
        "type": "float",
        "description": "Nucleus sampling probability; default 0.95. Must be positive.",
        "required": false,
        "max": 1,
        "default": 0.95
      },
      {
        "name": "min_p",
        "type": "float",
        "description": "Minimum probability relative to the most likely token; default 0.05.",
        "required": false,
        "min": 0.0,
        "max": 1.0,
        "default": 0.05
      },
      {
        "name": "repetition_penalty",
        "type": "float",
        "description": "Autoregressive repetition penalty; default 1.2.",
        "required": false,
        "min": 1.0,
        "default": 1.2
      },
      {
        "name": "text_chunk_size",
        "type": "int",
        "description": "Maximum codepoints per progressively emitted streaming segment; default 160.",
        "required": false,
        "min": 1,
        "default": 160
      },
      {
        "name": "text_chunk_mode",
        "type": "enum",
        "description": "Framework text chunking mode used by streaming synthesis.",
        "values": [
          "default",
          "tag_aware",
          "japanese",
          "endline"
        ],
        "required": false,
        "default": "default"
      },
      {
        "name": "seed",
        "type": "int",
        "description": "Sampling seed; omitted requests choose a random seed.",
        "required": false,
        "min": 0
      }
    ],
    "session": [
      {
        "name": "reference_cache_slots",
        "type": "int",
        "description": "Encoded reference-voice cache slots; default 1. Set to 0 to disable reuse.",
        "required": false,
        "min": 0,
        "default": 1
      },
      {
        "name": "backbone_weight_type",
        "type": "enum",
        "preset": "weight_type_full",
        "required": false,
        "default": "native",
        "description": "Storage type for the Qwen2 language-model weights."
      },
      {
        "name": "linear_weight_type",
        "type": "enum",
        "preset": "weight_type_full",
        "required": false,
        "default": "native",
        "description": "Storage type for non-convolutional speaker, processor, and decoder weights."
      },
      {
        "name": "conv_weight_type",
        "type": "enum",
        "required": false,
        "default": "f32",
        "description": "Storage type for speaker and processor convolution weights. The DAC transposed-convolution decoder remains F32 for CUDA compatibility. Applies to the speaker encoder/acoustic processor; waveform decoder remains f32.",
        "values": [
          "native",
          "f32",
          "f16",
          "bf16"
        ]
      }
    ],
    "load": []
  },
  "packages": [],
  "dependencies": [],
  "ui": {
    "tags": [
      "TTS",
      "Clone"
    ],
    "docs": [
      "docs/models/mira_tts.md"
    ]
  },
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "config": "model:config.json",
        "tokenizer_config": "model:tokenizer_config.json",
        "tokenizer_json": "model:tokenizer.json"
      },
      "tensors": {
        "language_model": {
          "source": "weights:",
          "prefix": "language_model"
        },
        "speaker_encoder": {
          "source": "weights:",
          "prefix": "speaker_encoder"
        },
        "processor": {
          "source": "weights:",
          "prefix": "processor"
        },
        "decoder": {
          "source": "weights:",
          "prefix": "decoder"
        },
        "upsampler": {
          "source": "weights:",
          "prefix": "upsampler"
        }
      }
    },
    {
      "format": "safetensors",
      "roots": {
        "model": "."
      },
      "files": {
        "config": "model:config.json",
        "tokenizer_config": "model:tokenizer_config.json",
        "tokenizer_json": "model:tokenizer.json"
      },
      "tensors": {
        "language_model": "model:language_model.safetensors",
        "speaker_encoder": "model:speaker_encoder.safetensors",
        "processor": "model:processor.safetensors",
        "decoder": "model:decoder.safetensors",
        "upsampler": "model:upsampler.safetensors"
      }
    }
  ]
}