{
  "schema_version": "transcribe-intake-v1",
  "family": "parakeet",
  "hf_repo": "nvidia/parakeet-rnnt-0.6b",
  "hf_revision": "c0c1f09fdc3f18b0b2ddbeafd5d6684f1b38078f",
  "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. n_mels=80 expected (matches FastConformer-Large convention pre-XL/v2)."
    },
    "tokenizer_config": {
      "kind": "hf_file",
      "path": "tokenizer_config.json",
      "status": "missing",
      "detail": "SentencePiece BPE inside .nemo, vocab_size=1024, no PnC."
    },
    "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-rnnt-0.6b",
      "status": "found",
      "detail": "Lowercase English ASR, no PnC, per-set WER published"
    },
    "reference_modeling_code": {
      "kind": "reference_code",
      "path": "scripts/convert-parakeet.py, src/arch/parakeet/",
      "status": "found",
      "detail": "Same encoder dims as TDT-0.6B-v2 (24L/d=1024/8h); only the joint differs (no duration head)."
    }
  },
  "variants": [
    {
      "name": "parakeet-rnnt-0.6b",
      "memory_gb": 2.5,
      "files": ["parakeet-rnnt-0.6b.nemo"]
    }
  ],
  "config": {
    "architecture_candidates": ["encoder-transducer"],
    "key_fields": {
      "architectures": ["EncDecRNNTBPEModel"],
      "model_type": "parakeet_rnnt",
      "encoder.type": "ConformerEncoder (FastConformer-Large)",
      "encoder.n_layers": 24,
      "encoder.d_model": 1024,
      "encoder.n_heads": 8,
      "encoder.ff_expansion_factor": 4,
      "encoder.conv_kernel_size": 9,
      "encoder.subsampling": "dw_striding",
      "encoder.subsampling_factor": 8,
      "encoder.subsampling_conv_channels": 256,
      "encoder.self_attention_model": "rel_pos",
      "decoder.type": "RNNTDecoder (predictor)",
      "decoder.pred_rnn_layers": 2,
      "decoder.pred_hidden": 640,
      "joint.type": "RNNTJoint",
      "joint.durations": null,
      "joint.num_extra_outputs": 0,
      "tokenizer.vocab_size": 1024
    },
    "varying_across_variants": []
  },
  "dtype": {
    "expected": "float32",
    "source": "manual",
    "evidence": "NeMo .nemo, PyTorch state_dict at float32.",
    "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"],
    "streaming": false,
    "speaker_diarization": false
  },
  "upstream_benchmarks": [
    {
      "dataset": "LibriSpeech test-clean",
      "language": "en",
      "metric": "wer",
      "score": 1.63,
      "score_unit": "percent",
      "source": "https://huggingface.co/nvidia/parakeet-rnnt-0.6b",
      "notes": null
    },
    {
      "dataset": "LibriSpeech test-other",
      "language": "en",
      "metric": "wer",
      "score": 3.06,
      "score_unit": "percent",
      "source": "https://huggingface.co/nvidia/parakeet-rnnt-0.6b",
      "notes": null
    }
  ],
  "reference_framework": "nemo",
  "reference_rationale": "First-party NeMo framework; same argument as the rest of the family. Encoder is identical to parakeet-tdt-0.6b-v2 (FastConformer-Large 24L/d=1024); only the joint differs (RNN-T, no duration head). The cleanest reuse target in this batch.",
  "architecture_pattern": "encoder-transducer",
  "known_risks": [
    "Plain RNN-T joint (no duration head). The existing TDT joint code needs a config-driven branch to skip the duration head; otherwise reuses 1:1.",
    "Encoder dims expected identical to TDT-0.6B-v2 (FastConformer-Large 24L/1024d/8h). Confirm from .nemo at convert time before declaring full reuse.",
    "n_mels=80 (older FastConformer-Large) vs v2's 128. This is the clearest dim-shift risk in the port — a wrong default here silently degrades WER.",
    "Standard parakeet risks (dw_striding mask, rel_pos shift, preemph ordering, per-feature norm, blank id at vocab_size) apply identically.",
    "No PnC; direct WER eval against LibriSpeech."
  ],
  "intake_gaps": [
    {
      "field": "frontend.n_mels",
      "reason": "Stated as 80 here based on FastConformer-Large era convention; confirm by reading model.cfg.preprocessor.features at convert time. Wrong value silently degrades WER."
    },
    {
      "field": "tokenizer.vocab_sha256",
      "reason": "Computed at convert time."
    }
  ]
}