🌐🎙️ Shenava Koochik v1.0 — ONNX FP16

The browser deployment of Shenava Koochik v1.0. Shenava’s web client loads this export with onnxruntime-web, preferring WebGPU and falling back to WASM. Inference is local after the model is downloaded.

✨ At a glance | معرفی سریع

English فارسی
🌐 Role Shenava browser model مدل مرورگر شنوا
⚡ Runtime ONNX Runtime Web: WebGPU → WASM اجرای WebGPU با fallback به WASM
📦 Format FP16, embedded or external-data FP16 تک‌فایلی یا ONNX/data
🎧 Input 16 kHz audio → 80-bin log-mel صوت ۱۶ کیلوهرتز → log-mel هشتادباند
🔒 Privacy Local inference after download اجرای محلی پس از دانلود

📦 Choose one graph layout

Files Best use
*_embedded.onnx Single-file download; simplest browser/WASM integration.
*.onnx + *.onnx.data External-data pair; used by Shenava’s WebGPU and modern WASM paths.

Both layouts contain the same FP16 fixed-window CTC export (ONNX opset 18, IR 10). export_manifest.json records their hashes and provenance.

📐 Exact contract

  • Audio preprocessing: 16 kHz mono, 80-bin NeMo-compatible log-mel, 25 ms window, 10 ms hop, normalize=NA.
  • Input processed_signal: FP16 [batch, 80, 2005].
  • Input processed_signal_length: INT64 [batch].
  • Output logits: FP16 [batch, 252, 1025].
  • Output encoded_lengths: INT64 [batch].
  • CTC blank: token 1024; vocabulary and decoding rules are in tokens.json.

Use preprocessor.json and mel_filters_slaney_80x257.json; small feature-pipeline differences materially change transcripts. This graph is a fixed-window model, not the cache-aware streaming graph. Live captions require windowing/overlap in the host application. For stateful streaming, use the tract or iOS 15 export.

✅ Published parity

Against the FP32 NeMo source on 269 gold clips: 268/269 exact transcripts, 0.9997 argmax agreement, and 0.9998 character similarity.

🚀 Download

from huggingface_hub import snapshot_download

snapshot_download(
    "Reza2kn/Shenava-Koochik-v1.0-ONNX-fp16",
    local_dir="./shenava-koochik-onnx",
)

🇮🇷 خلاصهٔ فارسی

این نسخه برای اجرای محلی در مرورگر با ONNX Runtime Web ساخته شده است. فایل embedded ساده‌ترین دانلود تک‌فایلی است و جفت ONNX/data برای WebGPU و WASM جدید مناسب است. پیش‌پردازش دقیق sidecarها الزامی است؛ خود گراف پنجره‌ای است و state جریانی ندارد.

🌌 Explore Shenava-1

🧠 Koochik source · 🌐 Browser ONNX · 🧰 sherpa-onnx · 🍎 Core ML · 🦀 tract streaming

Apache-2.0. Numbers are emitted in spoken form; apply Persian ITN in the display layer if digits are required.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for PersianML/Shenava-Koochik-v1.0-ONNX-fp16

Quantized
(15)
this model

Datasets used to train PersianML/Shenava-Koochik-v1.0-ONNX-fp16