{
  "schema_version": "transcribe-intake-v1",
  "family": "moonshine",
  "hf_repo": "UsefulSensors/moonshine-base",
  "hf_revision": "7a73d8d55ac0ba2ef3ae761593f6784b51f96dcf",
  "sources": {
    "config": {
      "kind": "hf_file",
      "path": "config.json",
      "status": "found"
    },
    "preprocessor": {
      "kind": "hf_file",
      "path": "preprocessor_config.json",
      "status": "found"
    },
    "tokenizer_config": {
      "kind": "hf_file",
      "path": "tokenizer_config.json",
      "status": "missing",
      "detail": "no tokenizer_config.json on the repo; tokenizer.json + generation_config.json carry the special-token IDs"
    },
    "tokenizer_json": {
      "kind": "hf_file",
      "path": "tokenizer.json",
      "status": "found"
    },
    "generation_config": {
      "kind": "hf_file",
      "path": "generation_config.json",
      "status": "found"
    },
    "safetensors_metadata": {
      "kind": "hf_api",
      "path": "HfApi.get_safetensors_metadata",
      "status": "found",
      "detail": "header-only floating dtype distribution; no tensor payloads downloaded"
    },
    "model_card": {
      "kind": "hf_api",
      "path": "https://huggingface.co/UsefulSensors/moonshine-base",
      "status": "found",
      "detail": "model card text shared with moonshine-tiny (license MIT, English-only, paper link); per-model WER scores not present in card body — taken from the Moonshine paper / Open ASR Leaderboard"
    },
    "reference_modeling_code": {
      "kind": "reference_code",
      "path": "transformers.models.moonshine (MoonshineForConditionalGeneration, AutoProcessor wrapping Wav2Vec2FeatureExtractor + PreTrainedTokenizerFast); cross-check refs/mlx/mlx-audio/mlx_audio/stt/models/moonshine/",
      "status": "found",
      "detail": "Same reference framework as moonshine-tiny — first-party in transformers since 4.48, no trust_remote_code."
    }
  },
  "variants": [
    {
      "name": "moonshine-base",
      "memory_gb": null,
      "files": []
    }
  ],
  "config": {
    "architecture_candidates": [
      "encoder-decoder"
    ],
    "key_fields": {
      "architectures": [
        "MoonshineForConditionalGeneration"
      ],
      "model_type": "moonshine",
      "is_encoder_decoder": true,
      "hidden_size": 416,
      "intermediate_size": 1664,
      "encoder_num_hidden_layers": 8,
      "decoder_num_hidden_layers": 8,
      "encoder_num_attention_heads": 8,
      "decoder_num_attention_heads": 8,
      "encoder_num_key_value_heads": 8,
      "decoder_num_key_value_heads": 8,
      "encoder_hidden_act": "gelu",
      "decoder_hidden_act": "silu",
      "max_position_embeddings": 194,
      "partial_rotary_factor": 0.62,
      "rope_theta": 10000.0,
      "attention_bias": false,
      "pad_head_dim_to_multiple_of": 8,
      "vocab_size": 32768
    },
    "varying_across_variants": [
      "hidden_size",
      "intermediate_size",
      "encoder_num_hidden_layers",
      "decoder_num_hidden_layers",
      "partial_rotary_factor"
    ]
  },
  "dtype": {
    "expected": "float32",
    "source": "config",
    "evidence": "config torch_dtype=float32; safetensors header distribution F32=210 (100%)",
    "details": {
      "config_declared": "float32",
      "header_distribution": {
        "F32": 210
      }
    }
  },
  "frontend": {
    "sample_rate": 16000,
    "n_mels": 1,
    "hop_length": null,
    "fft_size": null,
    "window": null,
    "normalization": "none",
    "preemphasis": null,
    "dither": null,
    "center": null,
    "padding_mode": null,
    "mel_filterbank_norm": null
  },
  "tokenizer": {
    "type": "bpe",
    "vocab_size": 32768,
    "special_tokens": {
      "bos": 1,
      "eos": 2,
      "pad": 2,
      "decoder_start": 1
    },
    "has_language_tokens": false,
    "vocab_sha256": "167ef33e0aa23978fd1bfa643785768b75ad36ff247f06d7f3c5e4e506453a80"
  },
  "capabilities": {
    "languages": ["en"],
    "language_detection": false,
    "translation": false,
    "timestamps": [],
    "streaming": false,
    "speaker_diarization": false
  },
  "upstream_benchmarks": [
    {
      "dataset": "LibriSpeech test-clean",
      "language": "en",
      "metric": "wer",
      "score": 3.27,
      "score_unit": "percent",
      "source": "https://arxiv.org/abs/2410.15608",
      "notes": "Moonshine paper Table 2; also matches the Open ASR Leaderboard."
    },
    {
      "dataset": "LibriSpeech test-other",
      "language": "en",
      "metric": "wer",
      "score": 8.30,
      "score_unit": "percent",
      "source": "https://arxiv.org/abs/2410.15608",
      "notes": "Moonshine paper / Open ASR Leaderboard."
    },
    {
      "dataset": "TED-LIUM",
      "language": "en",
      "metric": "wer",
      "score": 4.74,
      "score_unit": "percent",
      "source": "https://arxiv.org/abs/2410.15608",
      "notes": "Open ASR Leaderboard (TED-LIUM 3)."
    },
    {
      "dataset": "Common Voice (mean)",
      "language": "en",
      "metric": "wer",
      "score": 8.25,
      "score_unit": "percent",
      "source": "https://arxiv.org/abs/2410.15608",
      "notes": "Open ASR Leaderboard (Common Voice 16.1 English)."
    },
    {
      "dataset": "GigaSpeech",
      "language": "en",
      "metric": "wer",
      "score": 12.40,
      "score_unit": "percent",
      "source": "https://arxiv.org/abs/2410.15608",
      "notes": "Open ASR Leaderboard."
    },
    {
      "dataset": "AMI",
      "language": "en",
      "metric": "wer",
      "score": 17.79,
      "score_unit": "percent",
      "source": "https://arxiv.org/abs/2410.15608",
      "notes": "Open ASR Leaderboard."
    },
    {
      "dataset": "Earnings22",
      "language": "en",
      "metric": "wer",
      "score": 12.85,
      "score_unit": "percent",
      "source": "https://arxiv.org/abs/2410.15608",
      "notes": "Open ASR Leaderboard."
    },
    {
      "dataset": "SPGISpeech",
      "language": "en",
      "metric": "wer",
      "score": 5.13,
      "score_unit": "percent",
      "source": "https://arxiv.org/abs/2410.15608",
      "notes": "Open ASR Leaderboard."
    },
    {
      "dataset": "VoxPopuli",
      "language": "en",
      "metric": "wer",
      "score": 12.20,
      "score_unit": "percent",
      "source": "https://arxiv.org/abs/2410.15608",
      "notes": "Open ASR Leaderboard."
    }
  ],
  "reference_framework": "transformers",
  "reference_rationale": "Same canonical reference as moonshine-tiny: the UsefulSensors model card demonstrates direct use of `MoonshineForConditionalGeneration` + `AutoProcessor` from mainline `transformers` (no `trust_remote_code`). The Moonshine classes have been first-party in transformers since 4.48 and carry the publisher-maintained generation config, tokenizer, and feature extractor — they are the canonical instrumentable reference. `refs/mlx/mlx-audio/mlx_audio/stt/models/moonshine/` is a useful clean Python cross-read but is not the published entrypoint. The dumper at scripts/dump_reference_moonshine_transformers.py already covers this path for moonshine-tiny and works for moonshine-base by repo-id alone.",
  "architecture_pattern": "encoder-decoder",
  "known_risks": [
    "All seven moonshine-tiny risks carry forward unchanged: novel raw-PCM conv frontend, variable-length input, partial RoPE, dual MLP shapes (encoder GELU vs decoder SwiGLU), tied word embeddings, dual KV-cache regime (cross-attn precomputed once vs self-attn per-step growth), hallucination throttling. The C++ port at src/arch/moonshine/ already handles all of these.",
    "Variant deltas vs tiny that the C++ port must keep correct: hidden_size 416 (vs 288), intermediate_size 1664 (vs 1152), encoder/decoder layers 8 (vs 6), partial_rotary_factor 0.62 (vs 0.9). All four are config-driven inputs to the existing weights.cpp/encoder.cpp/decoder.cpp graph — no structural difference is expected. If any was hardcoded against tiny's value, validate.py at Stage 4 will surface it.",
    "head_dim padding: moonshine-base nominal head_dim = 416/8 = 52 (vs 36 on tiny). With pad_head_dim_to_multiple_of=8, padded head_dim = 56 (vs 40 on tiny). FlashAttention/SDPA paths in C++ that hardcoded head_dim=40 to match the dk40_dv40 Metal kernel are a likely friction point. Confirm at Stage 4 whether the existing fa kernel selection is data-driven or hardcoded.",
    "Partial RoPE rotary dim: tiny rotates 32 of 36 head dims (int(36*0.9)=32, even-rounded). Base rotates int(52*0.62)=32 of 52 head dims, even-rounded = 32 — coincidentally the same rotary dim, but the unrotated tail is 20 dims (vs 4 on tiny). Cohere-style partial RoPE handling in the existing port should already split rotated/unrotated by data; verify."
  ],
  "intake_gaps": []
}