PersianML/Shenava-Koochik-Lite-v1.0

🪶🎙️ Shenava Koochik Lite v1.0

A LITEASR-compressed encoder for Shenava Koochik v1.0. Post-training low-rank factorization reduces the encoder from 108.9M to 85.4M parameters (21.6%) without retraining.

This repository is not a standalone ASR checkpoint. It contains a replacement encoder state dict and must be loaded on top of the base .nemo model; the decoder, CTC head, and tokenizer still come from Koochik.

✨ At a glance | معرفی سریع

	English	فارسی
🪶 Role	Compressed Koochik encoder	encoder فشرده‌شدهٔ کوچیک
📉 Reduction	108.9M → 85.4M encoder parameters	کاهش ۲۱٫۶ درصدی پارامترهای encoder
🧪 Method	Post-training LITEASR low-rank factorization	فشرده‌سازی low-rank بدون آموزش مجدد
🧩 Requirement	Base Koochik .nemo is required	فایل NeMo مدل اصلی الزامی است
⚠️ Scope	Not a standalone checkpoint	checkpoint مستقل نیست

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

📦 Files

• koochik_lite099_enc.pt: compressed FP32 encoder state dict.
• koochik_lite099_kmap.json: retained rank for each factorized layer.
• load_koochik_lite.py: reconstructs the low-rank modules and loads the state dict into the base model.

🚀 Load

from huggingface_hub import hf_hub_download, snapshot_download
from load_koochik_lite import load_koochik_lite

base = hf_hub_download(
    "Reza2kn/Shenava-Koochik-v1.0",
    "shenava-koochik-v1.0.nemo",
)
repo = snapshot_download("Reza2kn/Shenava-Koochik-Lite-v1.0")
model = load_koochik_lite(
    base,
    f"{repo}/koochik_lite099_enc.pt",
    f"{repo}/koochik_lite099_kmap.json",
)
print(model.transcribe(["speech.wav"])[0].text)

📊 Published trade-off

The release evaluated both greedy decoding and an optional Vosk-guided hotword beam. Lower is better.

Decode	golden-6669 keyword-band WER	FLEURS keyword-band WER	golden-6669 overall WER	FLEURS overall WER
Full Koochik, greedy	8.0	13.1	4.64	5.36
Koochik Lite, greedy	12.5	18.0	6.92	7.23
Koochik Lite + Vosk guide	6.4	11.7	5.30	5.31

Compression alone reduces quality; the Vosk-guided result requires a separate Vosk first pass plus hotword-aware pyctcdecode beam search. Do not compare the guided row to a greedy-only deployment as though they used the same runtime.

🇮🇷 خلاصهٔ فارسی

این مخزن یک مدل کامل و مستقل نیست؛ فقط encoder فشرده‌شده را نگه می‌دارد و برای اجرا به فایل NeMo مدل اصلی نیاز دارد. نسخهٔ greedy سبک‌تر است ولی افت دقت دارد؛ ردیف Vosk-guided به یک مرحلهٔ جداگانهٔ Vosk و beam search نیاز دارد.

🌌 Explore Shenava-1

🧠 Full Koochik · 🪶 Koochik Lite · ⚖️ Rizeh 32M · 🐣 Rizeh-Pizeh 6.9M

Apache-2.0. Compression method: LITEASR.