Shenava-Koochik-v1.0

🎙️ Shenava Koochik v1.0 · شنوا کوچیک

The 114M-parameter flagship of the Shenava-1 Persian ASR family. This is the FP32 NeMo source checkpoint for evaluation, fine-tuning, and export; use one of the deployment repositories below for browser, Core ML, sherpa-onnx, or tract applications.

✨ At a glance | معرفی سریع

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

🧠 Model

• Architecture: FastConformer Hybrid RNNT/CTC; the CTC head is used by the published deployment exports.
• Audio: mono, 16 kHz Persian speech.
• Encoder: d_model=512, 17 layers, 8x subsampling (about 80 ms per encoder step).
• Contexts: [70,13], [70,6], [70,1], and [70,0].
• Tokenizer: ve_tok_v4, SentencePiece BPE-1024 plus CTC blank.
• Output: Persian text. Numbers are emitted in spoken form; apply Persian inverse text normalization when digits are wanted.

The two .nemo filenames are compatibility aliases with identical content and SHA-256:

f7b5124a9fd2d50c15bf070abdc0f80ec2449d64948c5657adb79a7e91dd1d16

📊 Published evaluation

Decoded with context [70,13] and the double-benchmark ITN/Persian-digit normalization convention.

🚀 Load with NeMo

from nemo.collections.asr.models import ASRModel

model = ASRModel.restore_from("shenava-koochik-v1.0.nemo")
text = model.transcribe(["speech.wav"])[0].text
print(text)

🧩 Deployment variants

• ONNX FP16: fixed-window browser/WebGPU/WASM export used by Shenava.
• sherpa-onnx: offline cross-platform bundle.
• Core ML FP16: fixed-window ML Program package.
• Core ML iOS 15: cache-aware streaming NeuralNetwork model for older Apple OS versions.
• tract streaming and tract offline: pure-Rust runtime exports.

🇮🇷 خلاصهٔ فارسی

این مخزن نسخهٔ اصلی FP32 و NeMo مدل ۱۱۴ میلیون‌پارامتری «شنوا کوچیک» است. برای اجرای مرورگر، Core ML، sherpa-onnx یا tract از مخزن مخصوص همان قالب استفاده کنید. ورودی گفتار فارسی تک‌کانالهٔ ۱۶ کیلوهرتز و خروجی متن فارسی است.

🌌 Explore Shenava-1

🧠 Koochik 114M · ⚖️ Rizeh 32M · 🐣 Rizeh-Pizeh 6.9M · 🌐 Browser · 🍎 Apple · 🦀 Rust

Apache-2.0. ASR quality can vary with accent, noise, overlap, recording channel, and code-switching; review meaning-critical transcripts before relying on them.