{ "schema_version": "transcribe-intake-v1", "family": "parakeet", "hf_repo": "nvidia/parakeet-tdt-1.1b", "hf_revision": "53276c6469d1f17a1352e30c4d11be3d0d7e9575", "sources": { "config": { "kind": "hf_file", "path": "config.json", "status": "missing", "detail": "NeMo ships the model as a single .nemo tar archive containing model_config.yaml. config.json is not on the HF repo root; values below come from the .nemo archive consumed by NeMo's ASRModel.from_pretrained and from the parakeet family's existing converter/encoder code." }, "preprocessor": { "kind": "hf_file", "path": "preprocessor_config.json|feature_extractor_config.json", "status": "missing", "detail": "preprocessor lives inside the .nemo archive as model_config.yaml -> preprocessor (NeMo AudioToMelSpectrogramPreprocessor / FilterbankFeatures). Values below carried from the existing parakeet family ports; n_mels in particular must be confirmed from the .nemo at convert time (older 1.1B FastConformer-XL configs typically use 80 mels, newer use 128)." }, "tokenizer_config": { "kind": "hf_file", "path": "tokenizer_config.json", "status": "missing", "detail": "tokenizer is a SentencePiece BPE model embedded in the .nemo archive (tokenizer.model, tokenizer.vocab). vocab_size = 1024." }, "tokenizer_json": { "kind": "hf_file", "path": "tokenizer.json", "status": "missing", "detail": "no fast-tokenizer JSON; SentencePiece proto only" }, "generation_config": { "kind": "hf_file", "path": "generation_config.json", "status": "missing", "detail": "TDT decoding config (durations, max_symbols_per_step) lives in the .nemo archive under model.cfg.decoding" }, "safetensors_metadata": { "kind": "hf_api", "path": "HfApi.get_safetensors_metadata", "status": "missing", "detail": "weights are PyTorch state_dict inside the .nemo tarball, not safetensors; dtype filled manually from NeMo defaults" }, "model_card": { "kind": "hf_api", "path": "https://huggingface.co/nvidia/parakeet-tdt-1.1b", "status": "found", "detail": "language list, capability flags, and per-dataset WER scores read from the model card" }, "reference_modeling_code": { "kind": "reference_code", "path": "scripts/convert-parakeet.py, src/arch/parakeet/, refs/mlx/mlx-audio/mlx_audio/stt/models/parakeet/", "status": "found", "detail": "Existing parakeet TDT ports (v2/v3) are the structural reference. Converter and C++ code already cover this architecture pattern." } }, "variants": [ { "name": "parakeet-tdt-1.1b", "memory_gb": 4.5, "files": ["parakeet-tdt-1.1b.nemo"] } ], "config": { "architecture_candidates": ["encoder-transducer"], "key_fields": { "architectures": ["EncDecRNNTBPEModel"], "model_type": "parakeet_tdt", "encoder.type": "FastConformer (XL 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, "tokenizer.vocab_size": 1024 }, "varying_across_variants": [ "encoder.n_layers", "encoder.d_model", "encoder.feat_in" ] }, "dtype": { "expected": "float32", "source": "manual", "evidence": "NeMo .nemo archive ships the TDT model as float32 PyTorch state_dict (no safetensors). Project policy preserves storage dtype; converted GGUF will be parakeet-tdt-1.1b-F32.gguf, matching the existing TDT-0.6B ports.", "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": 1.39, "score_unit": "percent", "source": "https://huggingface.co/nvidia/parakeet-tdt-1.1b", "notes": null }, { "dataset": "LibriSpeech test-other", "language": "en", "metric": "wer", "score": 2.62, "score_unit": "percent", "source": "https://huggingface.co/nvidia/parakeet-tdt-1.1b", "notes": null }, { "dataset": "Open ASR Leaderboard (8-set average)", "language": "en", "metric": "wer", "score": null, "score_unit": "percent", "source": "https://huggingface.co/nvidia/parakeet-tdt-1.1b", "notes": "Per-set scores reported on card: AMI 15.90, Earnings22 14.65, GigaSpeech 9.55, SPGI 3.42, TEDLIUM-v3 3.56, VoxPopuli 5.48, CommonVoice 5.48. Aggregate not published; computed by Stage 7." } ], "reference_framework": "nemo", "reference_rationale": "Same as the existing parakeet-tdt-0.6b ports: NVIDIA NeMo is the first-party framework for the entire Parakeet family. The HF repo ships only a .nemo tar archive consumable by ASRModel.from_pretrained; there is no Transformers modeling class. NeMo is also the only framework that exposes the TDT joint / duration head numerics consistently. mlx-audio is a cross-check, not an authority.", "architecture_pattern": "encoder-transducer", "known_risks": [ "Encoder dimensions for the 1.1B variant (n_layers, d_model, n_heads, ff_expansion_factor, conv_kernel_size, feat_in, subsampling_conv_channels) are NOT yet locked in this intake — must be read from model_config.yaml inside the .nemo archive at convert time. The existing 0.6B port uses 24 layers / d_model=1024 / 8 heads; the 1.1B variant is likely XL-class FastConformer with different dims.", "TDT (Token-and-Duration Transducer) joint network: same risk as v2 — decoding loop must consume both token and duration heads per step; off-by-one in duration handling produces silent transcript drift, not a crash. Reuse the existing TDT joint code path.", "FastConformer encoder uses depth-wise striding subsampling (factor 8). Frame count must round consistently with NeMo's get_seq_len for the encoder mask to match.", "Relative-position attention (rel_pos) with untied biases. Reuse the existing parakeet rel-pos shift implementation.", "NeMo FilterbankFeatures applies preemph=0.97 BEFORE windowing/STFT. n_mels assumed=80 for this 1.1B variant; CONFIRM at convert time — older FastConformer-XL configs use 80, newer 0.6B v2/v3 use 128. Wrong n_mels silently degrades WER without changing shape.", "Per-feature normalization (mean/std across time, per mel band). Identical to v2.", "SentencePiece BPE tokenizer with appended at vocab_size (id 1024). Joint emits vocab+1 logits; id==1024 means blank. Identical to v2.", "Vocab is 1024 (English only, no PnC) — model card explicitly states 'lower case English alphabet'. No punctuation or capitalization to validate at Stage 7." ], "intake_gaps": [ { "field": "config.key_fields (encoder dims)", "reason": "encoder.n_layers, d_model, n_heads, ff_expansion_factor, conv_kernel_size, feat_in are not on the HF repo root and not in this intake. Read from .nemo model_config.yaml at convert time and update key_fields then." }, { "field": "frontend.n_mels", "reason": "Assumed 80 from FastConformer-XL convention; preprocessor_config.json is missing on HF (lives in .nemo). Must be cross-checked against model.cfg.preprocessor.features at convert time. If actual is 128, update before Stage 4." }, { "field": "tokenizer.vocab_sha256", "reason": "SentencePiece proto extracted from .nemo at convert time; SHA can be computed during conversion but not yet recorded here." }, { "field": "variants[].memory_gb", "reason": "Estimated from F32 weight-count rule of thumb (~4 bytes/param); not measured against actual peak inference memory." } ] }