Shenava-Koochik-1.0-CoreML-iOS15-fp16

🍎📱🎙️ Shenava Koochik 1.0 — Core ML streaming for iOS 15

Cache-aware FP16 Core ML NeuralNetwork export for Apple devices that cannot use the newer ML Program package. It uses Core ML specification version 5, has iOS 14/macOS 11 availability, and is designed to compile and run on iOS 15.

✨ At a glance | معرفی سریع

• Canonical repository: Reza2kn/Shenava-Koochik-1.0-CoreML-iOS15-fp16
• PersianML mirror: PersianML/Shenava-Koochik-1.0-CoreML-iOS15-fp16
• Parent: Reza2kn/Shenava-Koochik-v1.0

🌊 Streaming contract

Each call produces one CTC step from a 170 ms feature window and returns the caches for the next call.

Inputs:

• processed_signal: FP32 [1, 80, 17]
• cache_last_channel: FP32 [17, 1, 70, 512]
• cache_last_time: FP32 [17, 1, 512, 8]

Outputs:

• logits: FP32 [1, 1, 1025]
• cache_last_channel_next: FP32 [17, 1, 70, 512]
• cache_last_time_next: FP32 [17, 1, 512, 8]

The graph uses 9 frames of pre-encode cache plus an 8-frame chunk, 17 conformer layers, and a constant channel-cache length of 70. Initialize both caches to zero and thread each *_next output into the next call.

📦 Files

• *_ios15_fp16.mlmodel: 212 MB NeuralNetwork model.
• *_manifest.json: authoritative streaming shapes and export provenance.
• tokens.json, mel_filters_slaney_80x257.json, preprocessor.json: tokenizer/feature assets.
• export_koochik10_streaming_coreml.py: reproducible exporter.

For the streaming graph, treat shenava_koochik_1_0_ctc_streaming_att70_0_ios15_fp16_manifest.json and the export script as authoritative. The fixed-window fields retained in preprocessor.json describe the shared feature pipeline, not this graph’s 17-frame tensor contract.

Published FP16-versus-FP32 argmax agreement: 1.000.

🛠️ Compile for iOS 15

xcrun coremlcompiler compile \
  shenava_koochik_1_0_ctc_streaming_att70_0_ios15_fp16.mlmodel \
  ./CompiledModels \
  --deployment-target 15.0 \
  --platform ios

🇮🇷 خلاصهٔ فارسی

این نسخه برای اجرای جریانی روی iOS 15 و دستگاه‌های قدیمی‌تر اپل ساخته شده است. در هر فراخوانی ۱۷۰ میلی‌ثانیه پردازش می‌شود و cacheهای خروجی باید به فراخوانی بعدی داده شوند. قرارداد دقیق tensorها در فایل manifest قرار دارد.

🌌 Explore Shenava-1

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

Apache-2.0. Output numbers are spoken-form Persian unless the host applies ITN.