Modelpublic

FluidInference/parakeet-ultra-coreml

sourceHugging Facecc-by-4.0updated 1d agoView on Hugging Face

3likes23downloads

Model Card

parakeet-ultra-coreml

Core ML build of moondream/parakeet-ultra, a full-precision post-training of nvidia/parakeet-tdt-0.6b-v3. Same architecture, 25 languages, tokenizer and output conventions as v3; drop-in for the v3 component contract. Encoder is int8 linear per-channel (595 MB); decoder and joint are re-exported from the Ultra checkpoint; preprocessor and vocabulary are the v3 ones.

Usage (FluidAudio ≥ next release)

swift

let models = try await AsrModels.downloadAndLoad(version: .ultra)
let asr = AsrManager()
try await asr.initialize(models: models)
let result = try await asr.transcribe(audioFileURL)

bash

swift run fluidaudiocli transcribe audio.wav --model-version ultra

Accuracy (FluidAudio, M-series Mac, full corpora, corpus-level WER)

	v3 (Core ML)	ultra (this repo)
LibriSpeech test-clean (2620 files)	2.27 %	2.13 %
LibriSpeech test-other (2939 files)	4.12 %	3.81 %
FLEURS, 24 languages × 100 samples, mean	14.81 %	11.67 %

Ultra wins all 24 FLEURS languages. The int8 encoder scores the same WER as an fp16 export (2.13 % / 3.79 %). Speed is within a few percent of v3 run back to back on the same machine (test-clean 126.7× vs 128.6× on the ANE).

Files

File	Notes
`Encoder.mlmodelc`	595 MB, int8 linear per-channel, iOS 17+/macOS 14+, fixed 15 s window (`mel [1,128,1501]` → `encoder [1,1024,188]`)
`Decoder.mlmodelc`	RNNT prediction net, fp16, iOS 17+
`JointDecisionv3.mlmodelc`	single-step joint + top-K 64, fp16, iOS 17+
`Preprocessor.mlmodelc`	v3 mel front-end
`parakeet_vocab.json`, `parakeet_v3_vocab.json`	v3 vocabulary

Conversion recipe: FluidInference/mobius models/stt/parakeet-ultra/coreml. License CC-BY-4.0, as the upstream checkpoint.