{ "schema_version": "transcribe-intake-v1", "family": "parakeet", "hf_repo": "nvidia/parakeet-tdt-0.6b-v2", "hf_revision": "1b149a3589351c96ddb101709fe7dd9c7069572f", "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 scripts/convert-parakeet.py:281+ which reads model.cfg.preprocessor / model.cfg.encoder / model.cfg.decoder." }, "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 cross-checked against scripts/convert-parakeet.py and refs/mlx/mlx-audio/mlx_audio/stt/models/parakeet/audio.py." }, "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). Read by NeMo's tokenizer factory; surfaced via scripts/convert-parakeet.py:167+." }, "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, etc.) 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 and convert-parakeet.py output" }, "model_card": { "kind": "hf_api", "path": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2", "status": "found", "detail": "languages, capability flags, and Open ASR Leaderboard score read from the model card" }, "reference_modeling_code": { "kind": "reference_code", "path": "refs/mlx/mlx-audio/mlx_audio/stt/models/parakeet/, refs/models/parakeet/parakeet-mlx/parakeet_mlx/parakeet.py, transcribe.cpp/scripts/convert-parakeet.py, transcribe.cpp/src/arch/parakeet/", "status": "found", "detail": "C++ reference port and converter are the authoritative source for shape/frontend facts; MLX implementations are cross-checks" } }, "variants": [ { "name": "parakeet-tdt-0.6b-v2", "memory_gb": 2.5, "files": ["parakeet-tdt-0.6b-v2.nemo"] }, { "name": "parakeet-tdt-0.6b-v3", "memory_gb": 2.5, "files": ["model.safetensors", "config.json", "tokenizer.json"] } ], "config": { "architecture_candidates": ["encoder-transducer"], "key_fields": { "architectures": ["EncDecRNNTBPEModel"], "model_type": "parakeet_tdt", "encoder.type": "ConformerEncoder (FastConformer)", "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", "encoder.feat_in": 128, "decoder.type": "RNNTDecoder (predictor)", "decoder.pred_rnn_layers": 2, "decoder.pred_hidden": 640, "decoder.vocab_size": 1024, "joint.type": "TDTJoint", "joint.durations": [0, 1, 2, 3, 4], "joint.num_extra_outputs": 5 }, "varying_across_variants": [ "decoder.vocab_size", "joint.vocabulary_size", "tokenizer.vocab_size" ] }, "dtype": { "expected": "float32", "source": "manual", "evidence": "NeMo .nemo archive ships the TDT model as float32 PyTorch state_dict (no safetensors). scripts/convert-parakeet.py preserves storage dtype per project policy 2; converted GGUF is parakeet-tdt-0.6b-v2-F32.gguf. Cross-checked: refs/mlx/mlx-audio's parakeet weights are float32.", "details": { "config_declared": null, "header_distribution": {} }, "expected_f32_tensors": [] }, "frontend": { "sample_rate": 16000, "n_mels": 128, "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": "Open ASR Leaderboard (average across 8 English test sets)", "language": "en", "metric": "wer", "score": 6.05, "score_unit": "percent", "source": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2", "notes": "Publisher-claimed average WER across LibriSpeech test-clean/other, AMI, Earnings22, GigaSpeech, SPGISpeech, TEDLIUM, VoxPopuli at the time the v2 model was released." } ], "reference_framework": "nemo", "reference_rationale": "NeMo (`nvidia/NeMo`) is NVIDIA's first-party training and inference framework for the Parakeet family. The HF repo ships only a `.nemo` tar archive consumable by `nemo.collections.asr.models.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 and parakeet-mlx are independent re-implementations and are useful only as cross-checks.", "architecture_pattern": "encoder-transducer", "known_risks": [ "TDT (Token-and-Duration Transducer) joint network. Standard RNN-T joint emits one token per frame; TDT additionally emits a duration symbol that advances the encoder pointer by 0..N frames. Decoding loop must consume both heads per step; off-by-one in duration handling produces silent transcript drift, not a crash.", "FastConformer encoder uses depth-wise striding subsampling (factor 8) — three conv stacks with stride-2 in time. Frame count must round consistently with NeMo's `get_seq_len` for the encoder mask to match; mismatches surface as edge-of-utterance hallucinations.", "Relative-position attention (`rel_pos`) with untied biases. The pos-bias is a learned vector added to attention scores after a relative-shift trick. ggml has no first-class op; the existing port materialises the shift host-side per layer.", "NeMo FilterbankFeatures applies preemph=0.97 BEFORE windowing/STFT. Skipping this silently shifts the spectral tilt and degrades WER without changing shape. Cross-check via tensor dump comparison after the mel stage.", "Per-feature normalization (mean/std across time, per mel band). Distinct from per-utterance global-mean-var; mismatch with reference is a common bring-up bug.", "SentencePiece BPE tokenizer with `` appended at vocab_size (id 1024 for v2, 8192 for v3). The blank id is OUTSIDE the SentencePiece vocab — the joint emits vocab+1 logits and id==vocab_size means blank.", "Streaming/chunked inference is supported by the architecture (att_context_size) but not by the first port; out of scope for v2 bring-up." ], "intake_gaps": [ { "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 GGUF size; not measured against actual peak inference memory." } ] }