Modelpublic

k3rb3l/wav2vec2-xls-r-300m-en-atc-uwb-atcc-coreml

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes3downloads

Model Card

wav2vec2-xls-r-300m-en-atc-uwb-atcc — Core ML

`Jzuluaga/wav2vec2-xls-r-300m-en-atc-uwb-atcc` converted to Core ML for on-device inference on iPhone/iPad (iOS 16+, FP16 mlprogram).

Usage

—Input audio: raw mono 16 kHz Float32, shape (1, samples), 1–60 s (flexible). Feature-extractor normalization is baked into the graph — feed raw samples.
—Output logits: CTC logits (1, frames, vocab). Greedy-decode: argmax over vocab, collapse repeats, drop the [PAD] blank (id 28), map ids via vocab.json, | (id 0) → space.

Chunk audio longer than 60 s (e.g. 10 s windows with 2 s stride).

Conversion & validation

Converted with coremltools from the original PyTorch weights (traced, FP16). Greedy CTC transcripts of the validation clip match the PyTorch baseline: match: True, max |logit Δ| = 0.1379.

Note: the original repo decodes with a KenLM language model (pyctcdecode). That LM is not part of this package — on-device decoding is greedy CTC, which scores slightly lower than the LM-rescored path.

Conversion script: `gmet/tools/convert_wav2vec2_coreml.py`