Restore Shenava emoji styling and bilingual visual hierarchy
Browse files
README.md
CHANGED
|
@@ -31,14 +31,24 @@ datasets:
|
|
| 31 |
- Reza2kn/fleurs-fa-benchmark
|
| 32 |
---
|
| 33 |
|
| 34 |
-
# Shenava Rizeh v1.0
|
| 35 |
|
| 36 |
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](https://huggingface.co/Reza2kn/Shenava-Koochik-v1.0) teacher. This repository contains the FP32 NeMo source checkpoint; published deployment formats live in separate repositories.
|
| 37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
- Canonical repository: [`Reza2kn/Shenava-Rizeh-v1.0`](https://huggingface.co/Reza2kn/Shenava-Rizeh-v1.0)
|
| 39 |
- PersianML mirror: [`PersianML/Shenava-Rizeh-v1.0`](https://huggingface.co/PersianML/Shenava-Rizeh-v1.0)
|
| 40 |
|
| 41 |
-
## Model contract
|
| 42 |
|
| 43 |
- Audio: mono, 16 kHz Persian speech.
|
| 44 |
- Encoder: `d_model=256`, 16 layers, 8x subsampling.
|
|
@@ -47,7 +57,7 @@ The 32M-parameter middle tier of Shenava-1: a Persian FastConformer Hybrid RNNT/
|
|
| 47 |
- Tokenizer: ve_tok_v4, SentencePiece BPE-1024 plus blank.
|
| 48 |
- Output: Persian text; numbers are spoken-form unless the display layer applies ITN.
|
| 49 |
|
| 50 |
-
## Published evaluation
|
| 51 |
|
| 52 |
Decoded with context `[70,13]` and the double-benchmark ITN/Persian-digit normalization convention.
|
| 53 |
|
|
@@ -56,7 +66,7 @@ Decoded with context `[70,13]` and the double-benchmark ITN/Persian-digit normal
|
|
| 56 |
| visualears-golden-6669 | 12.11% | 3.94% |
|
| 57 |
| FLEURS-fa | 14.45% | 5.10% |
|
| 58 |
|
| 59 |
-
## Load with NeMo
|
| 60 |
|
| 61 |
```python
|
| 62 |
from nemo.collections.asr.models import ASRModel
|
|
@@ -67,8 +77,12 @@ print(model.transcribe(["speech.wav"])[0].text)
|
|
| 67 |
|
| 68 |
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.
|
| 69 |
|
| 70 |
-
## فارسی
|
| 71 |
|
| 72 |
«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.
|
| 73 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
Apache-2.0. Accuracy varies with accent, noise, overlap, recording channel, and code-switching.
|
|
|
|
| 31 |
- Reza2kn/fleurs-fa-benchmark
|
| 32 |
---
|
| 33 |
|
| 34 |
+
# ⚖️🎙️ Shenava Rizeh v1.0 · شنوا ریزه
|
| 35 |
|
| 36 |
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](https://huggingface.co/Reza2kn/Shenava-Koochik-v1.0) teacher. This repository contains the FP32 NeMo source checkpoint; published deployment formats live in separate repositories.
|
| 37 |
|
| 38 |
+
## ✨ At a glance | معرفی سریع
|
| 39 |
+
|
| 40 |
+
| | English | فارسی |
|
| 41 |
+
|---|---|---|
|
| 42 |
+
| ⚖️ Role | Balanced 32M middle tier | مدل متعادل میانی با ۳۲M پارامتر |
|
| 43 |
+
| 🧠 Lineage | Distilled from 114M Koochik | تقطیرشده از کوچیک ۱۱۴M |
|
| 44 |
+
| 📦 Format | FP32 NeMo source | checkpoint اصلی FP32 و NeMo |
|
| 45 |
+
| 🎧 Input | 16 kHz mono Persian speech | گفتار فارسی تککانالهٔ ۱۶ کیلوهرتز |
|
| 46 |
+
| 🎯 Best for | Accuracy/footprint balance | تعادل دقت و اندازه |
|
| 47 |
+
|
| 48 |
- Canonical repository: [`Reza2kn/Shenava-Rizeh-v1.0`](https://huggingface.co/Reza2kn/Shenava-Rizeh-v1.0)
|
| 49 |
- PersianML mirror: [`PersianML/Shenava-Rizeh-v1.0`](https://huggingface.co/PersianML/Shenava-Rizeh-v1.0)
|
| 50 |
|
| 51 |
+
## 🧠 Model contract
|
| 52 |
|
| 53 |
- Audio: mono, 16 kHz Persian speech.
|
| 54 |
- Encoder: `d_model=256`, 16 layers, 8x subsampling.
|
|
|
|
| 57 |
- Tokenizer: ve_tok_v4, SentencePiece BPE-1024 plus blank.
|
| 58 |
- Output: Persian text; numbers are spoken-form unless the display layer applies ITN.
|
| 59 |
|
| 60 |
+
## 📊 Published evaluation
|
| 61 |
|
| 62 |
Decoded with context `[70,13]` and the double-benchmark ITN/Persian-digit normalization convention.
|
| 63 |
|
|
|
|
| 66 |
| visualears-golden-6669 | 12.11% | 3.94% |
|
| 67 |
| FLEURS-fa | 14.45% | 5.10% |
|
| 68 |
|
| 69 |
+
## 🚀 Load with NeMo
|
| 70 |
|
| 71 |
```python
|
| 72 |
from nemo.collections.asr.models import ASRModel
|
|
|
|
| 77 |
|
| 78 |
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.
|
| 79 |
|
| 80 |
+
## 🇮🇷 خلاصهٔ فارسی
|
| 81 |
|
| 82 |
«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.
|
| 83 |
|
| 84 |
+
## 🌌 Explore Shenava-1
|
| 85 |
+
|
| 86 |
+
[🧠 Koochik 114M](https://huggingface.co/Reza2kn/Shenava-Koochik-v1.0) · **⚖️ Rizeh 32M** · [🐣 Rizeh-Pizeh 6.9M](https://huggingface.co/Reza2kn/Shenava-Rizeh-Pizeh-v1.0)
|
| 87 |
+
|
| 88 |
Apache-2.0. Accuracy varies with accent, noise, overlap, recording channel, and code-switching.
|