{
  "schema_version": "transcribe-intake-v1",
  "family": "granite",
  "hf_repo": "ibm-granite/granite-speech-4.1-2b-plus",
  "hf_revision": "edd3bf54fbb06d8e263aa0c1939321d67b073f86",
  "sources": {
    "config": {
      "kind": "hf_file",
      "path": "config.json",
      "status": "found"
    },
    "preprocessor": {
      "kind": "hf_file",
      "path": "preprocessor_config.json|feature_extractor_config.json",
      "status": "accepted_gap",
      "detail": "Not shipped in the repo. The granite_speech_plus processor is registered in HEAD transformers and reuses GraniteSpeechFeatureExtractor; frontend values mirrored from the base 4.1-2b variant. Oracle (porting-2) must confirm by running the HEAD-transformers processor end-to-end."
    },
    "tokenizer_config": {
      "kind": "hf_file",
      "path": "tokenizer_config.json",
      "status": "found"
    },
    "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"
    },
    "reference_modeling_code": {
      "kind": "reference_code",
      "path": "transformers.models.granite_speech_plus (GraniteSpeechPlusForConditionalGeneration)",
      "status": "found",
      "detail": "Requires transformers HEAD (config pins 5.6.0.dev0; PyPI release expected at >=5.8). granite_speech_plus is a separate model_type from granite_speech and will not load under older transformers."
    },
    "model_card": {
      "kind": "hf_file",
      "path": "README.md",
      "status": "found",
      "detail": "Capabilities, languages, WER table, speaker-attribution and word-timestamp output protocols sourced here"
    }
  },
  "variants": [
    {
      "name": "granite-speech-4.1-2b-plus",
      "memory_gb": null,
      "files": []
    }
  ],
  "config": {
    "architecture_candidates": [
      "audio-llm"
    ],
    "key_fields": {
      "architectures": [
        "GraniteSpeechPlusForConditionalGeneration"
      ],
      "model_type": "granite_speech_plus",
      "audio_token_index": 100352,
      "downsample_rate": 5,
      "window_size": 15,
      "has_lora_adapter": false,
      "encoder.model_type": "granite_speech_plus_encoder",
      "encoder.num_layers": 16,
      "encoder.hidden_dim": 1024,
      "encoder.num_heads": 8,
      "encoder.dim_head": 128,
      "encoder.input_dim": 160,
      "encoder.output_dim": 348,
      "encoder.conv_kernel_size": 15,
      "encoder.conv_expansion_factor": 2,
      "encoder.feedforward_mult": 4,
      "encoder.max_pos_emb": 512,
      "encoder.context_size": 200,
      "encoder.cat_hidden_layers": [3],
      "projector.model_type": "blip_2_qformer",
      "projector.hidden_size": 1024,
      "projector.intermediate_size": 4096,
      "projector.num_attention_heads": 16,
      "projector.num_hidden_layers": 2,
      "projector.encoder_hidden_size": 2048,
      "projector.cross_attention_frequency": 1,
      "projector.position_embedding_type": "absolute",
      "projector.layer_norm_eps": 1e-12,
      "projector.vocab_size": 30522,
      "text.model_type": "granite",
      "text.hidden_size": 2048,
      "text.intermediate_size": 4096,
      "text.num_hidden_layers": 40,
      "text.num_attention_heads": 16,
      "text.num_key_value_heads": 4,
      "text.vocab_size": 100353,
      "text.max_position_embeddings": 4096,
      "text.rope_theta": 10000,
      "text.rms_norm_eps": 1e-05,
      "text.hidden_act": "silu",
      "text.embedding_multiplier": 12,
      "text.logits_scaling": 8,
      "text.attention_multiplier": 0.0078125,
      "text.residual_multiplier": 0.22,
      "text.tie_word_embeddings": true
    },
    "varying_across_variants": [
      "encoder_config.cat_hidden_layers",
      "projector_config.encoder_hidden_size",
      "model_type",
      "architectures",
      "transformers_version",
      "text_config.tie_word_embeddings"
    ]
  },
  "dtype": {
    "expected": "bfloat16",
    "source": "weights_header",
    "evidence": "Top-level config.dtype=bfloat16 and safetensors header BF16=937 (all floating tensors) agree. text_config.dtype=float32 is vestigial and inconsistent with both the top-level declaration and the actual weight storage. Treat top-level + header as authoritative.",
    "details": {
      "config_declared": "bfloat16",
      "header_distribution": {
        "BF16": 937,
        "I64": 16
      }
    },
    "expected_f32_tensors": []
  },
  "frontend": {
    "sample_rate": 16000,
    "n_mels": 80,
    "hop_length": 160,
    "fft_size": 512,
    "window": "hann_periodic",
    "normalization": "none",
    "preemphasis": null,
    "dither": 0.0,
    "center": true,
    "padding_mode": "reflect",
    "mel_filterbank_norm": "htk"
  },
  "tokenizer": {
    "type": "bpe",
    "vocab_size": 100353,
    "special_tokens": {
      "bos": 100257,
      "eos": 100257,
      "pad": 100256,
      "audio": 100352
    },
    "has_language_tokens": true,
    "vocab_sha256": "cd9faddbc5d020a41fb6dfa6924c090bf987ede9092bc088f6a7c66aef416af2"
  },
  "capabilities": {
    "languages": ["en", "fr", "de", "es", "pt"],
    "language_detection": false,
    "translation": false,
    "timestamps": ["word"],
    "streaming": false,
    "speaker_diarization": true
  },
  "upstream_benchmarks": [
    {
      "dataset": "LibriSpeech test-clean (Open ASR Leaderboard)",
      "language": "en",
      "metric": "wer",
      "score": 1.44,
      "score_unit": "percent",
      "source": "model card",
      "notes": "Ref-dtype C++ port must score <= 1.45 to pass porting-7-wer."
    },
    {
      "dataset": "LibriSpeech test-other (Open ASR Leaderboard)",
      "language": "en",
      "metric": "wer",
      "score": 3.06,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "TED-LIUM v3 (Open ASR Leaderboard)",
      "language": "en",
      "metric": "wer",
      "score": 3.89,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "SPGISpeech (Open ASR Leaderboard)",
      "language": "en",
      "metric": "wer",
      "score": 3.72,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "VoxPopuli (Open ASR Leaderboard)",
      "language": "en",
      "metric": "wer",
      "score": 5.9,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "AMI (Open ASR Leaderboard)",
      "language": "en",
      "metric": "wer",
      "score": 8.63,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "Earnings22 (Open ASR Leaderboard)",
      "language": "en",
      "metric": "wer",
      "score": 8.68,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "GigaSpeech (Open ASR Leaderboard)",
      "language": "en",
      "metric": "wer",
      "score": 10.38,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "FISHER (diarization WDER)",
      "language": "en",
      "metric": "other",
      "score": 0.9,
      "score_unit": "percent",
      "source": "model card",
      "notes": "Word-level diarization error rate (lower is better). Diarization scoring is not part of porting-7-wer."
    },
    {
      "dataset": "CALLHOME (diarization WDER)",
      "language": "en",
      "metric": "other",
      "score": 2.2,
      "score_unit": "percent",
      "source": "model card"
    }
  ],
  "reference_framework": "transformers",
  "reference_rationale": "Mainline transformers carries granite_speech_plus as of the HEAD branch (config pins 5.6.0.dev0; the model card directs users to install transformers from source >=5.8). The class GraniteSpeechPlusForConditionalGeneration is a separate registration from granite_speech; older transformers cannot load this checkpoint. No author repo is involved. We pin a HEAD-transformers SHA in scripts/envs/granite/pyproject.toml.",
  "architecture_pattern": "audio-llm",
  "known_risks": [
    "Encoder forward path differs from base granite_speech: with cat_hidden_layers=[3], the encoder hidden state at layer index 3 is concatenated along the feature dimension with the final-layer output before being fed into the projector. The projector's encoder_hidden_size=2048 (vs 1024 in the base) reflects this concatenation. Stage 4 must collect and concat the mid-layer hidden state correctly; missing this silently doubles the feature mismatch under the Q-Former.",
    "Output protocol embeds speaker tags ([Speaker N]:) and word-level timestamps ([T:N], N in centiseconds mod 1000) as text tokens. There are no structured fields. The C++ runtime emits a single text stream; downstream consumers parse the brackets. Stage 4 should add a parser-side test that exercises the text protocol; the model_card defines the exact format.",
    "preprocessor_config.json is missing from the repo. Frontend values above mirror the base 4.1-2b variant (GraniteSpeechFeatureExtractor defaults: 16kHz, n_fft=512, n_mels=80, hop=160, win=400). Oracle (porting-2) must confirm by loading the HEAD-transformers processor and dumping the realized mel features against a known wave.",
    "text_config.tie_word_embeddings=true here, while the base 4.1-2b and 4.0-1b variants have it false. If a single C++ loader serves the whole family, this distinction must be data-driven (e.g. via a GGUF KV pair) and must wire up llama.cpp-style TENSOR_DUPLICATED fallback for the lm_head when tied.",
    "Loses Japanese support vs base 4.1-2b: ja is absent from the supported language list. The C++ runtime must surface this difference in capability metadata; per-variant capability rows belong in the family doc's Capability Validation table.",
    "transformers version requirement is unstable (5.6.0.dev0 in the config, >=5.8 in the model card). Reference env should pin a specific commit SHA and the family doc must record both the pinned transformers SHA and the model card's recommended minimum at the time of port.",
    "All risks from the base granite_speech intake (Conformer block-attention, Shaw relpos, BLIP-2 Q-Former, Granite LLM scalar multipliers, GQA 16/4, audio token injection at id 100352, frontend torchaudio defaults) apply here unchanged."
  ],
  "intake_gaps": [
    {
      "field": "frontend",
      "reason": "preprocessor_config.json is not shipped. Values mirrored from base 4.1-2b. Oracle (porting-2) must confirm against HEAD transformers' granite_speech_plus processor."
    },
    {
      "field": "frontend.window/center/padding_mode/mel_filterbank_norm",
      "reason": "Same caveat as base granite_speech variants: torchaudio defaults assumed; oracle confirms."
    }
  ]
}