Shenava-Koochik-v1.0-CoreML-fp16

🍎🎙️ Shenava Koochik v1.0 — Core ML FP16

Fixed-window FP16 Core ML ML Program export of Shenava Koochik v1.0. The published package targets macOS 13 and is intended for modern Apple deployment where a complete audio window is processed at once.

✨ At a glance | معرفی سریع

• Canonical repository: Reza2kn/Shenava-Koochik-v1.0-CoreML-fp16
• PersianML mirror: PersianML/Shenava-Koochik-v1.0-CoreML-fp16

🔎 What this is—and is not

• shenava-koochik-v1.0_ctc_fixed2005_len_fp16.mlpackage is a fixed-window ML Program package.
• Input features: processed_signal [1, 80, 2005] plus the valid feature length.
• Output: CTC logits with a 1,025-token vocabulary.
• Published FP16 parity on 269 gold clips: 261/269 exact transcripts, 0.9980 argmax agreement, 0.9972 character similarity.
• Package size: about 209 MB.

This is not the legacy iOS 15 model and it is not cache-aware streaming. For 170 ms stateful streaming on older Apple devices, use Shenava-Koochik-1.0-CoreML-iOS15-fp16.

🧩 Host requirements

The repository contains the Core ML package only. Reuse the 16 kHz/80-mel preprocessing and ve_tok_v4 decoding sidecars from the ONNX FP16 repository. Preserve normalize=NA, the valid-length input, CTC blank id 1024, and SentencePiece ▁ word boundaries.

🛠️ Compile

xcrun coremlcompiler compile \
  shenava-koochik-v1.0_ctc_fixed2005_len_fp16.mlpackage \
  ./CompiledModels

🇮🇷 خلاصهٔ فارسی

این نسخه یک ML Program با پنجرهٔ ثابت برای macOS 13 و دستگاه‌های جدید اپل است. این مدل state جریانی ندارد. برای اجرای جریانی ۱۷۰ میلی‌ثانیه‌ای روی iOS 15 از مخزن CoreML-iOS15 استفاده کنید.

🌌 Explore Shenava-1

🧠 Koochik source · 🍎 Modern Core ML · 📱 iOS 15 streaming · 🌐 Browser ONNX · 🦀 Rust

Apache-2.0. Numbers are emitted in spoken form; apply Persian ITN at display time when digits are wanted.