Modelpublic

NeoRoth/parakeet-fr-qat-dropout01-20260919-coreml-int8

sourceHugging Facecc-by-4.0updated 6d agoView on Hugging Face

0likes33downloads

Model Card

Parakeet French QAT Dropout 0.1 Core ML INT8

French-specialized Parakeet TDT v3 bundle for the Oriloq macOS application. It is distributed as a Core ML encoder plus the native predictor and joint network required by Oriloq.

Intended use

Use this model from Oriloq on Apple Silicon Macs running macOS 15 or later. The bundle is not a generic Transformers or NeMo checkpoint, and it is designed and supported for the Oriloq integration.

Architecture

—Core ML compiled Conformer encoder with per-channel INT8 weight quantization and FP16 compute.
—Core ML configuration requests CPU and Neural Engine execution.
—The native LSTM predictor and joint network remain on CPU.
—The package includes fixed encoder functions for 501, 1501, and 3001 MEL frames.

Evaluation

The published artifact is the epoch 14 checkpoint selected on Oriloq's local French validation set: 3.00% normalized WER and 6.74% strict WER. On the FLEURS French development split, it measured 6.97% WER, versus 4.97% for the base model. These results describe different evaluation distributions and do not establish a general-purpose French quality claim.

The current Oriloq Core ML integration test loads this exact bundle and verifies a non-empty transcription containing "bonjour".

Included runtime files

—config.json
—encoder.mlmodelc/
—predictor_joint.safetensors
—SentencePiece and tokenizer files

No training audio, transcripts, or dataset material are included.

License and attribution

This derivative is released under CC BY 4.0. It is based on NVIDIA Parakeet TDT 0.6B v3.