Modelpublic

soloish90/parakeet-unified-en-0.6b-onnx-coreml

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes10downloads

Model Card

Parakeet Unified 0.6B — ONNX + CoreML streaming runtime graphs

Streaming-inference export of nvidia/parakeet-unified-en-0.6b (FastConformer-RNNT, unified punctuation/casing), split into the three graphs used by Babble's in-process runtime. Attention context [70, 7, 0] (chunked streaming with right context) is baked in.

File	Stage	Format	Runs on
`featurizer.onnx` (+`.data`)	audio → 128-mel features	ONNX fp16	CPU
`encoder.mlpackage/`	features → encoder states	CoreML mlprogram fp16, static 617-frame window, masks as inputs	Apple GPU (macOS 13+)
`encoder.onnx`	features → encoder states	ONNX fp16, same masked static contract	CPU / non-Apple GPU
`decoder_joint.onnx` (+`.data`)	one RNN-T label-loop step	ONNX fp16	CPU
`tokenizer_decode.json`	id → piece decode table	JSON	—

The two encoders are interchangeable: identical inputs (features [1,128,617] fp32 zero-padded, features_length, pad_mask [1,78], att_mask [1,78,78] floats, 1.0 = pad/blocked) and outputs (encoded [1,1024,78], encoded_length). A platform downloads only the encoder it uses.

Provenance & validation

Exported by the conversion scripts in Babble's babble-onnx runtime (TorchScript trace of the NeMo checkpoint; ONNX opset 17; CoreML via coremltools 9, fp16 compute). Every artifact is gated against the original NeMo implementation: token-exact greedy decode on a 24-fixture parity suite, a 38-clip library sweep, and a 42-clip batch-replay corpus (~80 minutes of speech, zero word differences; the only divergences ever observed are single punctuation marks on measured sub-0.01-logit decisions).

License

CC-BY-4.0, inherited from the NVIDIA base model. Weights are a direct format conversion of the original checkpoint; attribution to NVIDIA per the base model card.