LEMAS-Project/LEMAS-Edit

1{2  "root": "pretrained_models",3  "structure": {4    "files": [5      {6        "path": "pretrained_models/denoiser_model.onnx",7        "info": "Audio denoising model (DeepFilterNet)"8      }9    ],10    "ckpts": {11      "autoregressive": {12        "path": "pretrained_models/ckpts/autoregressive",13        "description": "Autoregressive neural codec editing model and its dependencies",14        "files": [15          {16            "path": "pretrained_models/ckpts/autoregressive/config.json",17            "info": "Model architecture and hyperparameter configuration for the autoregressive editing model"18          },19          {20            "path": "pretrained_models/ckpts/autoregressive/multilingual_330M.pth",21            "info": "Main weights of the multilingual autoregressive codec editing model (~330M parameters), used when Edit Model=autoregressive"22          },23          {24            "path": "pretrained_models/ckpts/autoregressive/encodec_4cb2048_giga.th",25            "info": "Neural codec encoder (AudioTokenizer) with 4 codebooks × 2048 codewords, used to decode discrete codes into 16kHz waveforms"26          },27          {28            "path": "pretrained_models/ckpts/autoregressive/dac_SR_8codes_2048_hop960_speech.pth",29            "info": "DAC super-resolution model (AudioSR), used to reconstruct audio from 16kHz codec to 48kHz or target sample rate for the AR backend"30          }31        ]32      },33      "multilingual_grl": {34        "path": "pretrained_models/ckpts/multilingual_grl",35        "description": "Main LEMAS-TTS model (multilingual_grl)",36        "files": [37          {38            "path": "pretrained_models/ckpts/multilingual_grl/multilingual_grl.safetensors",39            "info": "Unconditional/text-conditioned non-autoregressive model weights with GRL, supporting multilingual TTS and editing (default Edit Model in Gradio)"40          }41        ]42      },43      "multilingual_prosody": {44        "path": "pretrained_models/ckpts/multilingual_prosody",45        "description": "non-autoregressive variant with an additional prosody encoder",46        "files": [47          {48            "path": "pretrained_models/ckpts/multilingual_prosody/multilingual_prosody.safetensors",49            "info": "Multilingual model weights with global prosody conditioning enabled (selectable in the model menu)"50          }51        ]52      },53      "prosody_encoder": {54        "path": "pretrained_models/ckpts/prosody_encoder",55        "description": "Pretssel / UnitY2-style prosody encoder used by the non-autoregressive prosody backend",56        "files": [57          {58            "path": "pretrained_models/ckpts/prosody_encoder/pretssel_cfg.json",59            "info": "Architecture configuration for the Pretssel/UnitY2 prosody encoder (dimensions, number of layers, etc.)"60          },61          {62            "path": "pretrained_models/ckpts/prosody_encoder/prosody_encoder_UnitY2.pt",63            "info": "Prosody encoder weights, used to extract global prosody embeddings from reference audio"64          }65        ]66      },67      "vocos-mel-24khz": {68        "path": "pretrained_models/ckpts/vocos-mel-24khz",69        "description": "Vocos neural vocoder (mel → 24kHz waveform)",70        "files": [71          {72            "path": "pretrained_models/ckpts/vocos-mel-24khz/config.yaml",73            "info": "Vocos vocoder configuration defining mel feature dimensions and network architecture"74          },75          {76            "path": "pretrained_models/ckpts/vocos-mel-24khz/pytorch_model.bin",77            "info": "Main Vocos vocoder weights, used to decode mel features in the CFM backend"78          },79          {80            "path": "pretrained_models/ckpts/vocos-mel-24khz/README.md",81            "info": "Documentation for the vocoder (origin, usage, and notes)"82          }83        ]84      }85    },86    "data": {87      "multilingual_grl": {88        "path": "pretrained_models/data/multilingual_grl",89        "description": "Text vocabulary for the multilingual_grl model",90        "files": [91          {92            "path": "pretrained_models/data/multilingual_grl/vocab.txt",93            "info": "phone vocabulary corresponding to the text embeddings of the multilingual_grl model"94          }95        ]96      },97      "multilingual_prosody": {98        "path": "pretrained_models/data/multilingual_prosody",99        "description": "Text vocabulary for the multilingual_prosody model",100        "files": [101          {102            "path": "pretrained_models/data/multilingual_prosody/vocab.txt",103            "info": "Shared text vocabulary used by the multilingual_prosody model"104          }105        ]106      }107    },108    "demos": {109      "root": "pretrained_models/demos",110      "files": [111        {112          "path": "pretrained_models/demos/test.wav",113          "info": "Simple test audio used for quick validation of gradio script"114        }115      ],116      "lemas_edit_test": {117        "path": "pretrained_models/demos/lemas_edit_test",118        "description": "Audio samples and alignment annotations for LEMAS-Edit demos",119        "subdirs": {120          "vocals": {121            "path": "pretrained_models/demos/lemas_edit_test/vocals",122            "files": [123              {124                "path": "pretrained_models/demos/lemas_edit_test/vocals/en_AUD0000000214_S0001522.mp3",125                "info": "English demo audio used for AR/NAR editing examples"126              },127              {128                "path": "pretrained_models/demos/lemas_edit_test/vocals/zh_emilia_zh_0008385782.mp3",129                "info": "Chinese demo audio used for multilingual editing examples"130              }131            ]132          },133          "align": {134            "path": "pretrained_models/demos/lemas_edit_test/align",135            "files": [136              {137                "path": "pretrained_models/demos/lemas_edit_test/align/en_AUD0000000214_S0001522.json",138                "info": "MMS alignment JSON for the English demo, including intervals, words, and modified_index used for editing"139              },140              {141                "path": "pretrained_models/demos/lemas_edit_test/align/zh_emilia_zh_0008385782.json",142                "info": "MMS alignment JSON for the Chinese demo"143              }144            ]145          }146        }147      }148    },149    "uvr5": {150      "path": "pretrained_models/uvr5",151      "description": "Kim Vocal UVR5 models and configurations (used for denoising)",152      "files": [153        {154          "path": "pretrained_models/uvr5/Kim_Vocal_1.onnx",155          "info": "Main UVR5 model (ONNX format) for vocal/accompaniment separation and denoising"156        },157        {158          "path": "pretrained_models/uvr5/MDX-Net-Kim-Vocal1.json",159          "info": "UVR5 model architecture and inference configuration (channels, frame length, etc.)"160        },161        {162          "path": "pretrained_models/uvr5/model_data.json",163          "info": "UVR5 metadata including presets, model list, and default parameters"164        },165        {166          "path": "pretrained_models/uvr5/model_name_mapper.json",167          "info": "Mapping from internal UVR5 model names to human-readable names for frontend selection"168        }169      ]170    },171    "whisperx": {172      "path": "pretrained_models/whisperx",173      "description": "WhisperX VAD and segmentation model assets",174      "files": [175        {176          "path": "pretrained_models/whisperx/whisperx-vad-segmentation.bin",177          "info": "WhisperX voice activity detection (VAD) weights used for long-audio segmentation and ASR alignment assistance"178        },179        {180          "path": "pretrained_models/whisperx/whisperx-vad-segmentation.bak",181          "info": "Backup or legacy version of the VAD model, kept for safety"182        }183      ]184    }185  }186}187