Modelpublic

Reza2kn/Shenava-Rizeh-v1.0

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

2likes208downloads

Model Card

⚖️🎙️ Shenava Rizeh v1.0 · شنوا ریزه

The 32M-parameter middle tier of Shenava-1: a Persian FastConformer Hybrid RNNT/CTC model distilled with logit and feature knowledge distillation from the 114M Koochik teacher. This repository contains the FP32 NeMo source checkpoint; published deployment formats live in separate repositories.

✨ At a glance | معرفی سریع

	English	فارسی
⚖️ Role	Balanced 32M middle tier	مدل متعادل میانی با ۳۲M پارامتر
🧠 Lineage	Distilled from 114M Koochik	تقطیرشده از کوچیک ۱۱۴M
📦 Format	FP32 NeMo source	checkpoint اصلی FP32 و NeMo
🎧 Input	16 kHz mono Persian speech	گفتار فارسی تک‌کانالهٔ ۱۶ کیلوهرتز
🎯 Best for	Accuracy/footprint balance	تعادل دقت و اندازه

—Canonical repository: `Reza2kn/Shenava-Rizeh-v1.0`
—PersianML mirror: `PersianML/Shenava-Rizeh-v1.0`

🧠 Model contract

—Audio: mono, 16 kHz Persian speech.
—Encoder: d_model=256, 16 layers, 8x subsampling.
—Contexts: [70,13], [70,6], [70,1], and [70,0].
—Deployed head: CTC.
—Tokenizer: vetokv4, SentencePiece BPE-1024 plus blank.
—Output: Persian text; numbers are spoken-form unless the display layer applies ITN.

📊 Published evaluation

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

Set	WER	CER
visualears-golden-6669	12.11%	3.94%
FLEURS-fa	14.45%	5.10%

🚀 Load with NeMo

python

from nemo.collections.asr.models import ASRModel

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

Choose Rizeh when Koochik’s accuracy/size trade-off is too heavy but the 6.9M Rizeh-Pizeh model is too small for the required accuracy.

🇮🇷 خلاصهٔ فارسی

«شنوا ریزه» مدل میانی ۳۲ میلیون‌پارامتری خانوادهٔ Shenava-1 است. این مخزن checkpoint اصلی FP32 و NeMo را نگه می‌دارد و برای ارزیابی، fine-tune یا تبدیل به قالب‌های اجرایی مناسب است.

🌌 Explore Shenava-1

🧠 Koochik 114M · ⚖️ Rizeh 32M · 🐣 Rizeh-Pizeh 6.9M

Apache-2.0. Accuracy varies with accent, noise, overlap, recording channel, and code-switching.