fp16 CoreML ML Program of the CTC head of Reza2kn/shenava-fa-fastconformer-115m, targeting the Apple Neural Engine (iOS 16+ / macOS). Native on-device path for VisualEars.

Input: processed_signal [1, 80, 2005] — 80-dim log-mel features (fixed 20.05 s window).
Outputs: CTC logits [1, 251, 1025] + encoded length. Greedy-CTC decode; truncate to ceil(real_mel_frames / 8).
Golden6669 WER: 7.92% (CTC, official Persian normalizer); verified faithful (8.55% on a 40-clip sanity).
Speed: the ANE decodes the 20.05 s window in ~10–14 ms (~1400–1900× real time).

Fully offline, on-device. Part of the VisualEars project (SLT 2026).

Reza2kn/shenava-fa-fastconformer-115m-coreml-fp16

Shenava v4 — FastConformer-115M CTC — fp16 CoreML (.mlpackage, Apple Neural Engine)

Reza2kn/
shenava-fa-fastconformer-115m-coreml-fp16