{
  "schema_version": "transcribe-intake-v1",
  "family": "parakeet",
  "hf_repo": "nvidia/parakeet-tdt_ctc-110m",
  "hf_revision": "431a349f3051ab85c22b9b7a2741b5fe77065665",
  "sources": {
    "config": {
      "kind": "hf_file",
      "path": "config.json",
      "status": "missing",
      "detail": "NeMo .nemo archive only."
    },
    "preprocessor": {
      "kind": "hf_file",
      "path": "preprocessor_config.json",
      "status": "missing",
      "detail": "Inside .nemo as model.cfg.preprocessor. Small FastConformer typically uses n_mels=80."
    },
    "tokenizer_config": {
      "kind": "hf_file",
      "path": "tokenizer_config.json",
      "status": "missing",
      "detail": "SentencePiece BPE inside .nemo, vocab_size=1024 with PnC chars in vocab."
    },
    "tokenizer_json": {
      "kind": "hf_file",
      "path": "tokenizer.json",
      "status": "missing"
    },
    "generation_config": {
      "kind": "hf_file",
      "path": "generation_config.json",
      "status": "missing"
    },
    "safetensors_metadata": {
      "kind": "hf_api",
      "path": "HfApi.get_safetensors_metadata",
      "status": "missing",
      "detail": "PyTorch state_dict inside .nemo"
    },
    "model_card": {
      "kind": "hf_api",
      "path": "https://huggingface.co/nvidia/parakeet-tdt_ctc-110m",
      "status": "found",
      "detail": "PnC support, dual TDT/CTC decoding, per-set WER. Smallest variant in the family — useful as fast development workhorse."
    },
    "reference_modeling_code": {
      "kind": "reference_code",
      "path": "scripts/convert-parakeet.py, src/arch/parakeet/",
      "status": "found",
      "detail": "Existing TDT path applies; encoder dims will be smaller (S-class FastConformer)."
    }
  },
  "variants": [
    {
      "name": "parakeet-tdt_ctc-110m",
      "memory_gb": 0.5,
      "files": ["parakeet-tdt_ctc-110m.nemo"]
    }
  ],
  "config": {
    "architecture_candidates": ["encoder-transducer"],
    "key_fields": {
      "architectures": ["EncDecHybridRNNTCTCBPEModel"],
      "model_type": "parakeet_tdt_ctc",
      "encoder.type": "FastConformer (S/Small config)",
      "encoder.subsampling": "dw_striding",
      "encoder.subsampling_factor": 8,
      "encoder.self_attention_model": "rel_pos",
      "decoder.type": "RNNTDecoder (predictor)",
      "joint.type": "TDTJoint",
      "joint.durations": [0, 1, 2, 3, 4],
      "joint.num_extra_outputs": 5,
      "ctc_decoder.type": "ConvASRDecoder",
      "tokenizer.vocab_size": 1024,
      "tokenizer.has_punctuation_capitalization": true
    },
    "varying_across_variants": [
      "encoder.n_layers",
      "encoder.d_model"
    ]
  },
  "dtype": {
    "expected": "float32",
    "source": "manual",
    "evidence": "NeMo .nemo, PyTorch state_dict at float32 by convention.",
    "details": {
      "config_declared": null,
      "header_distribution": {}
    },
    "expected_f32_tensors": []
  },
  "frontend": {
    "sample_rate": 16000,
    "n_mels": 80,
    "hop_length": 160,
    "fft_size": 512,
    "window": "hann_periodic",
    "normalization": "per_feature",
    "preemphasis": 0.97,
    "dither": 1e-05,
    "center": true,
    "padding_mode": "reflect",
    "mel_filterbank_norm": "slaney"
  },
  "tokenizer": {
    "type": "sentencepiece",
    "vocab_size": 1024,
    "special_tokens": {
      "blank": 1024
    },
    "has_language_tokens": false,
    "vocab_sha256": null
  },
  "capabilities": {
    "languages": ["en"],
    "language_detection": false,
    "translation": false,
    "timestamps": ["token", "word", "segment"],
    "streaming": false,
    "speaker_diarization": false
  },
  "upstream_benchmarks": [
    {
      "dataset": "LibriSpeech test-clean",
      "language": "en",
      "metric": "wer",
      "score": 2.4,
      "score_unit": "percent",
      "source": "https://huggingface.co/nvidia/parakeet-tdt_ctc-110m",
      "notes": "WER computed against PnC-stripped, lower-cased reference."
    },
    {
      "dataset": "LibriSpeech test-other",
      "language": "en",
      "metric": "wer",
      "score": 5.2,
      "score_unit": "percent",
      "source": "https://huggingface.co/nvidia/parakeet-tdt_ctc-110m",
      "notes": null
    }
  ],
  "reference_framework": "nemo",
  "reference_rationale": "Same NeMo argument as the rest of the family. The 110M is the smallest variant and a natural fast-iteration target for porting validation.",
  "architecture_pattern": "encoder-transducer",
  "known_risks": [
    "Smallest variant in the family — encoder dims significantly different from the 0.6B port (likely fewer layers, smaller d_model). Lock from .nemo model_config.yaml at convert time.",
    "Hybrid TDT + CTC head sharing the encoder; same disposition decision as tdt_ctc-1.1b — ship TDT only by default.",
    "PnC in vocab: WER eval requires lowercasing + punct-strip. Same as tdt_ctc-1.1b.",
    "Standard parakeet TDT risks (joint duration, dw_striding mask, rel_pos shift, preemph ordering, per-feature norm, blank id).",
    "n_mels likely 80 (Small FastConformer convention) but confirm at convert time."
  ],
  "intake_gaps": [
    {
      "field": "config.key_fields (encoder dims)",
      "reason": "Smaller variant — different layer/dim counts than 0.6B. Lock at convert time from model_config.yaml."
    },
    {
      "field": "frontend.n_mels",
      "reason": "Assumed 80; confirm from .nemo at convert time."
    },
    {
      "field": "ctc_head_disposition",
      "reason": "Same decision as tdt_ctc-1.1b: ship TDT only or expose both heads."
    },
    {
      "field": "tokenizer.vocab_sha256",
      "reason": "Computed at convert time."
    }
  ]
}