{ "schema_version": "transcribe-intake-v1", "family": "parakeet", "hf_repo": "nvidia/parakeet-unified-en-0.6b", "hf_revision": "d4ac9928f3bf238223ff0779c06b8149bf8ac4e1", "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 likely 128 (newer 0.6B FastConformer matching v2 era), but confirm at convert time." }, "tokenizer_config": { "kind": "hf_file", "path": "tokenizer_config.json", "status": "missing", "detail": "SentencePiece BPE inside .nemo. PnC chars present in vocab." }, "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-unified-en-0.6b", "status": "found", "detail": "Unified offline+streaming RNN-T with shared parameters. Trained jointly with mode-consistency regularization. PnC support. Streaming latency configurable 160ms-2080ms." }, "reference_modeling_code": { "kind": "reference_code", "path": "scripts/convert-parakeet.py, src/arch/parakeet/", "status": "found", "detail": "Encoder dims expected to match FastConformer-Large 0.6B family. Only the joint differs from TDT (RNN-T, no duration head). Streaming-specific weights (Dynamic Chunked Conv) shared with offline path per model card." } }, "variants": [ { "name": "parakeet-unified-en-0.6b", "memory_gb": 2.5, "files": ["parakeet-unified-en-0.6b.nemo"] } ], "config": { "architecture_candidates": ["encoder-transducer"], "key_fields": { "architectures": ["EncDecRNNTBPEModel"], "model_type": "parakeet_rnnt_unified", "encoder.type": "Unified FastConformer (offline + streaming, shared weights)", "encoder.n_layers": 24, "encoder.subsampling": "dw_striding (non-causal in offline mode)", "encoder.subsampling_factor": 8, "encoder.self_attention_model": "rel_pos with chunked masking option", "encoder.streaming.left_context": "5.6s (default)", "encoder.streaming.right_context_range": "0-2080ms in 80ms steps", "encoder.streaming.dynamic_chunked_conv": true, "decoder.type": "RNNTDecoder (predictor)", "joint.type": "RNNTJoint", "joint.durations": null, "joint.num_extra_outputs": 0, "tokenizer.vocab_size": 1024, "tokenizer.has_punctuation_capitalization": true }, "varying_across_variants": [] }, "dtype": { "expected": "float32", "source": "manual", "evidence": "NeMo .nemo, PyTorch state_dict at float32 by family convention.", "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"], "streaming": true, "speaker_diarization": false }, "upstream_benchmarks": [ { "dataset": "LibriSpeech test-clean (offline)", "language": "en", "metric": "wer", "score": 1.63, "score_unit": "percent", "source": "https://huggingface.co/nvidia/parakeet-unified-en-0.6b", "notes": "Offline / full-context inference. WER scored against PnC-stripped, lower-cased reference." }, { "dataset": "LibriSpeech test-other (offline)", "language": "en", "metric": "wer", "score": 3.11, "score_unit": "percent", "source": "https://huggingface.co/nvidia/parakeet-unified-en-0.6b", "notes": "Offline / full-context inference." }, { "dataset": "Open ASR Leaderboard (8-set average, offline)", "language": "en", "metric": "wer", "score": 5.91, "score_unit": "percent", "source": "https://huggingface.co/nvidia/parakeet-unified-en-0.6b", "notes": "Offline. Streaming variants on the same set: 6.29 at 1.12s latency, 8.44 at 160ms latency. v1 transcribe.cpp port targets the offline number only." } ], "reference_framework": "nemo", "reference_rationale": "NVIDIA NeMo first-party framework. The unified offline+streaming training recipe is NeMo-specific (mode-consistency regularization loss + Dynamic Chunked Convolutions); only NeMo can authoritatively dump per-mode encoder intermediates.", "architecture_pattern": "encoder-transducer", "known_risks": [ "STREAMING IS PUNTED FOR v1 PORT. The model card states all encoder, predictor, and joint parameters are shared between offline and streaming modes — mode is selected at inference via attention masks and Dynamic Chunked Convolutions. The transcribe.cpp v1 port targets OFFLINE MODE ONLY (full-context attention, non-causal convolutions). Streaming can be added later without re-converting weights.", "Dynamic Chunked Convolutions: model card claims weights are vanilla 1-D depthwise (shared with offline). At Stage 3 convert time, verify there are no streaming-specific shape quirks (mask-aware param packing, cached state buffers stored as weights). Cheap to detect, expensive to debug if missed.", "Plain RNN-T joint (no duration head). Same as parakeet-rnnt-* — needs a config-driven branch in the existing TDT joint to skip duration.", "PnC in vocab: WER eval requires lowercasing + punct-stripping. Same as tdt_ctc-* variants.", "FastConformer-Large 0.6B encoder: dims expected identical to v2 (24L/d=1024/8h), but the unified variant adds chunked-attention parameters (left/right context buffers as state, not weights). Confirm at convert time.", "Best-in-class WER for the 0.6B size (1.63 / 3.11) — strong candidate to ship as the headline 0.6B English RNN-T." ], "intake_gaps": [ { "field": "config.key_fields (encoder dims, streaming params)", "reason": "Exact n_layers, d_model, n_heads not on the model card. Lock from .nemo model_config.yaml at convert time. Streaming-specific keys (chunk sizes, left/right context buffers) need cataloging at convert time even though offline-only is shipped." }, { "field": "frontend.n_mels", "reason": "Stated as 128 (newer 0.6B convention matching v2/v3) but unverified — the unified-en model is newer than the rnnt-0.6b but predates v2/v3 in some respects. Confirm from .nemo at convert time." }, { "field": "tokenizer.vocab_sha256", "reason": "Computed at convert time." }, { "field": "streaming_disposition", "reason": "v1 port ships offline-only. The streaming half of the model lives in the .nemo weights (as masks/state, per the card) but is not exercised. Document as a future Stage 4+ extension." } ] }