voice_audio_multi/model_specs/mms_forced_aligner.json

{
  "schema_version": 1,
  "family": "mms_forced_aligner",
  "display_name": "Meta MMS-300M Forced Aligner",
  "description": "Meta MMS-300M multilingual CTC forced aligner (Wav2Vec2ForCTC) that aligns an exact transcript to mono/stereo audio and returns word timestamps. The checkpoint covers 1,130 languages, but this port's native text normalization supports Dutch (nl/nld) and English (en/eng) Latin script only; text_normalization=pre_romanized accepts caller-supplied ASCII romanization, and unsupported non-Latin input is rejected. Checkpoint license: CC-BY-NC-4.0; GGUF conversion is local-only, no public redistribution is approved.",
  "category": "speech_analysis",
  "status": "community",
  "tasks": [
    "align"
  ],
  "modes": [
    "offline"
  ],
  "languages": [
    "multilingual"
  ],
  "capabilities": {
    "align": [
      "word_timestamps"
    ]
  },
  "dependencies": [],
  "options": {
    "request": [
      {
        "name": "language",
        "type": "string",
        "description": "Language code. Built-in normalization supports nl/nld and en/eng; pre_romanized accepts any nonempty code.",
        "required": true
      },
      {
        "name": "text_normalization",
        "type": "enum",
        "description": "Transcript normalization mode; latin handles Dutch/English Latin script natively, pre_romanized trusts caller-supplied ASCII romanization.",
        "values": [
          "latin",
          "pre_romanized"
        ],
        "required": false,
        "default": "latin"
      },
      {
        "name": "star_frequency",
        "type": "enum",
        "description": "Placement of the virtual <star> CTC target between words (segment) or at transcript edges (edges); matches the reference default segment.",
        "values": [
          "segment",
          "edges"
        ],
        "required": false,
        "default": "segment"
      },
      {
        "name": "merge_threshold_sec",
        "type": "float",
        "description": "Close gaps at or below this threshold, rounded to alignment frames, by moving the later word start; word entries remain separate. Default 0.0 disables gap closure.",
        "required": false,
        "min": 0.0,
        "default": 0.0
      },
      {
        "name": "return_timestamps",
        "type": "bool",
        "description": "Compatibility option; word timestamps are always returned, including when false.",
        "required": false,
        "default": true
      }
    ],
    "session": [
      {
        "name": "emission_window_sec",
        "type": "float",
        "description": "Center emission window length in seconds used to split long waveforms; must be > 0 and at most 3600. Rounded 16 kHz sample count must be a multiple of 320 (20 ms); default 30.",
        "required": false,
        "min": 0.02,
        "max": 3600,
        "default": 30.0
      },
      {
        "name": "emission_context_sec",
        "type": "float",
        "description": "Left/right context appended to each emission window in seconds; must be nonnegative, at most 3600, and below emission_window_sec. Rounded 16 kHz sample count must be a multiple of 320 (20 ms); default 2.",
        "required": false,
        "min": 0.0,
        "max": 3600,
        "default": 2.0
      },
      {
        "name": "max_alignment_cells",
        "type": "int",
        "description": "Hard cap on CTC DP cells (frames x states) allocated per request; alignment fails before allocation when exceeded; default 50000000.",
        "required": false,
        "min": 1,
        "default": 50000000
      },
      {
        "name": "max_target_tokens",
        "type": "int",
        "description": "Hard cap on flattened CTC target tokens per request; default 8192.",
        "required": false,
        "min": 1,
        "default": 8192
      },
      {
        "name": "weight_type",
        "type": "enum",
        "description": "Weight storage type; default native (tensors kept as stored in the checkpoint).",
        "preset": "weight_type_conv",
        "required": false,
        "default": "native"
      }
    ],
    "load": []
  },
  "runtime": {
    "tags": [
      "gguf"
    ]
  },
  "ui": {
    "recommended_package": "mms_forced_aligner_300m_f16",
    "tags": [
      "Align",
      "GGUF"
    ],
    "docs": [
      "docs/models/mms_forced_aligner.md",
      "docs/speech_analysis.md",
      "docs/gguf.md"
    ]
  },
  "package_defaults": {
    "download": {
      "kind": "unsupported",
      "reason": "CC-BY-NC-4.0 checkpoint: convert locally with audiocpp_gguf; no public audio.cpp GGUF distribution is approved."
    }
  },
  "packages": [
    {
      "id": "mms_forced_aligner_300m_f16",
      "display_name": "Meta MMS-300M Forced Aligner F16 GGUF",
      "default": true,
      "format": "gguf",
      "precision": "f16",
      "target_directory": "MMS-Forced-Aligner-GGUF",
      "files": [
        "MMS-Forced-Aligner-GGUF/mms-forced-aligner-f16.gguf"
      ],
      "strip_prefix": "MMS-Forced-Aligner-GGUF"
    },
    {
      "id": "mms_forced_aligner_300m_safetensors",
      "display_name": "Meta MMS-300M Forced Aligner Safetensors",
      "format": "safetensors",
      "precision": "native",
      "target_directory": "mms-300m-1130-forced-aligner",
      "files": [
        "config.json",
        "model.safetensors",
        "special_tokens_map.json",
        "tokenizer_config.json",
        "vocab.json"
      ],
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "MahmoudAshraf/mms-300m-1130-forced-aligner",
        "revision": "49402e9577b1158620820667c218cd494cc44486",
        "gated": false
      }
    },
    {
      "id": "mms_forced_aligner_300m_q8_0",
      "display_name": "Meta MMS-300M Forced Aligner Q8_0 GGUF",
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "MMS-Forced-Aligner-GGUF",
      "files": [
        "MMS-Forced-Aligner-GGUF/mms-forced-aligner-q8_0.gguf"
      ],
      "strip_prefix": "MMS-Forced-Aligner-GGUF"
    }
  ],
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "config": "model:config.json",
        "vocab": "model:vocab.json"
      },
      "optional_files": {
        "special_tokens_map": "model:special_tokens_map.json",
        "tokenizer_config": "model:tokenizer_config.json",
        "preprocessor_config": "model:preprocessor_config.json"
      },
      "tensors": {
        "weights": "weights:"
      }
    },
    {
      "format": "safetensors",
      "roots": {
        "model": "."
      },
      "files": {
        "config": "model:config.json",
        "vocab": "model:vocab.json"
      },
      "optional_files": {
        "special_tokens_map": "model:special_tokens_map.json",
        "tokenizer_config": "model:tokenizer_config.json",
        "preprocessor_config": "model:preprocessor_config.json"
      },
      "tensors": {
        "weights": "model:model.safetensors"
      }
    }
  ]
}