Hugging Face's logo

PersianML
/
Shenava-Koochik-Lite-v1.0

Automatic Speech Recognition
NeMo
Persian
speech
persian
farsi
fastconformer
ctc
on-device
shenava
shenava-1
visualears
liteasr
compression
low-rank
dhh
Model card Files
xet
 Community
Reza2kn commited on
Commit
6a409d2
·
verified ·
1 Parent(s): 191c760

Restore Shenava emoji styling and bilingual visual hierarchy

Browse files
Files changed (1) hide show
  1. README.md +19 -5
README.md CHANGED
@@ -27,22 +27,32 @@ datasets:
27
  - Reza2kn/fleurs-fa-benchmark
28
  ---
29
 
30
- # Shenava Koochik Lite v1.0
31
 
32
  A LITEASR-compressed encoder for [Shenava Koochik v1.0](https://huggingface.co/Reza2kn/Shenava-Koochik-v1.0). Post-training low-rank factorization reduces the encoder from 108.9M to 85.4M parameters (21.6%) without retraining.
33
 
34
  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.
35
 
 
 
 
 
 
 
 
 
 
 
36
  - Canonical repository: [`Reza2kn/Shenava-Koochik-Lite-v1.0`](https://huggingface.co/Reza2kn/Shenava-Koochik-Lite-v1.0)
37
  - PersianML mirror: [`PersianML/Shenava-Koochik-Lite-v1.0`](https://huggingface.co/PersianML/Shenava-Koochik-Lite-v1.0)
38
 
39
- ## Files
40
 
41
  - `koochik_lite099_enc.pt`: compressed FP32 encoder state dict.
42
  - `koochik_lite099_kmap.json`: retained rank for each factorized layer.
43
  - `load_koochik_lite.py`: reconstructs the low-rank modules and loads the state dict into the base model.
44
 
45
- ## Load
46
 
47
  ```python
48
  from huggingface_hub import hf_hub_download, snapshot_download
@@ -61,7 +71,7 @@ model = load_koochik_lite(
61
  print(model.transcribe(["speech.wav"])[0].text)
62
  ```
63
 
64
- ## Published trade-off
65
 
66
  The release evaluated both greedy decoding and an optional Vosk-guided hotword beam. Lower is better.
67
 
@@ -73,8 +83,12 @@ The release evaluated both greedy decoding and an optional Vosk-guided hotword b
73
 
74
  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.
75
 
76
- ## فارسی
77
 
78
  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.
79
 
 
 
 
 
80
  Apache-2.0. Compression method: [LITEASR](https://arxiv.org/abs/2502.20583).
 
27
  - Reza2kn/fleurs-fa-benchmark
28
  ---
29
 
30
+ # 🪶🎙️ Shenava Koochik Lite v1.0
31
 
32
  A LITEASR-compressed encoder for [Shenava Koochik v1.0](https://huggingface.co/Reza2kn/Shenava-Koochik-v1.0). Post-training low-rank factorization reduces the encoder from 108.9M to 85.4M parameters (21.6%) without retraining.
33
 
34
  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.
35
 
36
+ ## ✨ At a glance | معرفی سریع
37
+
38
+ | | English | فارسی |
39
+ |---|---|---|
40
+ | 🪶 Role | Compressed Koochik encoder | encoder فشرده‌شدهٔ کوچیک |
41
+ | 📉 Reduction | 108.9M → 85.4M encoder parameters | کاهش ۲۱٫۶ درصدی پارامترهای encoder |
42
+ | 🧪 Method | Post-training LITEASR low-rank factorization | فشرده‌سازی low-rank بدون آموزش مجدد |
43
+ | 🧩 Requirement | Base Koochik `.nemo` is required | فایل NeMo مدل اصلی الزامی است |
44
+ | ⚠️ Scope | Not a standalone checkpoint | checkpoint مستقل نیست |
45
+
46
  - Canonical repository: [`Reza2kn/Shenava-Koochik-Lite-v1.0`](https://huggingface.co/Reza2kn/Shenava-Koochik-Lite-v1.0)
47
  - PersianML mirror: [`PersianML/Shenava-Koochik-Lite-v1.0`](https://huggingface.co/PersianML/Shenava-Koochik-Lite-v1.0)
48
 
49
+ ## 📦 Files
50
 
51
  - `koochik_lite099_enc.pt`: compressed FP32 encoder state dict.
52
  - `koochik_lite099_kmap.json`: retained rank for each factorized layer.
53
  - `load_koochik_lite.py`: reconstructs the low-rank modules and loads the state dict into the base model.
54
 
55
+ ## 🚀 Load
56
 
57
  ```python
58
  from huggingface_hub import hf_hub_download, snapshot_download
 
71
  print(model.transcribe(["speech.wav"])[0].text)
72
  ```
73
 
74
+ ## 📊 Published trade-off
75
 
76
  The release evaluated both greedy decoding and an optional Vosk-guided hotword beam. Lower is better.
77
 
 
83
 
84
  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.
85
 
86
+ ## 🇮🇷 خلاصهٔ فارسی
87
 
88
  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.
89
 
90
+ ## 🌌 Explore Shenava-1
91
+
92
+ [🧠 Full Koochik](https://huggingface.co/Reza2kn/Shenava-Koochik-v1.0) · **🪶 Koochik Lite** · [⚖️ Rizeh 32M](https://huggingface.co/Reza2kn/Shenava-Rizeh-v1.0) · [🐣 Rizeh-Pizeh 6.9M](https://huggingface.co/Reza2kn/Shenava-Rizeh-Pizeh-v1.0)
93
+
94
  Apache-2.0. Compression method: [LITEASR](https://arxiv.org/abs/2502.20583).