Restore Shenava emoji styling and bilingual visual hierarchy
Browse files
README.md
CHANGED
|
@@ -30,15 +30,25 @@ datasets:
|
|
| 30 |
- Reza2kn/fleurs-fa-benchmark
|
| 31 |
---
|
| 32 |
|
| 33 |
-
# Shenava Rizeh-Pizeh v1.0
|
| 34 |
|
| 35 |
The smallest Shenava-1 Persian ASR model: a 6.9M-parameter FastConformer distilled through the Koochik → Rizeh → Rizeh-Pizeh cascade. This repository contains the FP32 NeMo source checkpoint for evaluation, fine-tuning, and export.
|
| 36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
- Canonical repository: [`Reza2kn/Shenava-Rizeh-Pizeh-v1.0`](https://huggingface.co/Reza2kn/Shenava-Rizeh-Pizeh-v1.0)
|
| 38 |
- PersianML mirror: [`PersianML/Shenava-Rizeh-Pizeh-v1.0`](https://huggingface.co/PersianML/Shenava-Rizeh-Pizeh-v1.0)
|
| 39 |
- Teacher: [`Reza2kn/Shenava-Rizeh-v1.0`](https://huggingface.co/Reza2kn/Shenava-Rizeh-v1.0)
|
| 40 |
|
| 41 |
-
## Model contract
|
| 42 |
|
| 43 |
- Audio: mono, 16 kHz Persian speech.
|
| 44 |
- Encoder: `d_model=144`, 12 layers, 8x subsampling.
|
|
@@ -49,7 +59,7 @@ The smallest Shenava-1 Persian ASR model: a 6.9M-parameter FastConformer distill
|
|
| 49 |
|
| 50 |
The release reported real-time FP32 tract inference on a 2015 Cortex-A7 (RTF about 0.91). Treat that as a release-specific device measurement, not a universal latency guarantee.
|
| 51 |
|
| 52 |
-
## Published evaluation
|
| 53 |
|
| 54 |
Decoded with context `[70,13]` and the double-benchmark ITN/Persian-digit normalization convention.
|
| 55 |
|
|
@@ -58,7 +68,7 @@ Decoded with context `[70,13]` and the double-benchmark ITN/Persian-digit normal
|
|
| 58 |
| visualears-golden-6669 | 24.55% | 8.89% |
|
| 59 |
| FLEURS-fa | 26.95% | 10.22% |
|
| 60 |
|
| 61 |
-
## Load with NeMo
|
| 62 |
|
| 63 |
```python
|
| 64 |
from nemo.collections.asr.models import ASRModel
|
|
@@ -69,8 +79,12 @@ print(model.transcribe(["speech.wav"])[0].text)
|
|
| 69 |
|
| 70 |
Choose this model when footprint and low-end CPU viability matter more than the accuracy available from the 32M Rizeh or 114M Koochik checkpoints.
|
| 71 |
|
| 72 |
-
## فارسی
|
| 73 |
|
| 74 |
«شنوا ریزهپیزه» کوچکترین مدل خانواده است: ۶٫۹ میلیون پارامتر برای اجرای کمهزینه روی CPUهای ضعیف. این مخزن checkpoint اصلی FP32 و NeMo را نگه میدارد؛ اندازهٔ کم با افت دقت نسبت به ریزه و کوچیک همراه است.
|
| 75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
Apache-2.0. Accuracy varies with accent, noise, overlap, recording channel, and code-switching.
|
|
|
|
| 30 |
- Reza2kn/fleurs-fa-benchmark
|
| 31 |
---
|
| 32 |
|
| 33 |
+
# 🐣🎙️ Shenava Rizeh-Pizeh v1.0 · شنوا ریزهپیزه
|
| 34 |
|
| 35 |
The smallest Shenava-1 Persian ASR model: a 6.9M-parameter FastConformer distilled through the Koochik → Rizeh → Rizeh-Pizeh cascade. This repository contains the FP32 NeMo source checkpoint for evaluation, fine-tuning, and export.
|
| 36 |
|
| 37 |
+
## ✨ At a glance | معرفی سریع
|
| 38 |
+
|
| 39 |
+
| | English | فارسی |
|
| 40 |
+
|---|---|---|
|
| 41 |
+
| 🐣 Role | Smallest Shenava-1 model | کوچکترین مدل خانوادهٔ Shenava-1 |
|
| 42 |
+
| 🪶 Scale | 6.9M parameters | ۶٫۹ میلیون پارامتر |
|
| 43 |
+
| 📦 Format | FP32 NeMo source | checkpoint اصلی FP32 و NeMo |
|
| 44 |
+
| 🧠 Lineage | Koochik → Rizeh → Rizeh-Pizeh | زنجیرهٔ تقطیر کوچیک ← ریزه ← ریزهپیزه |
|
| 45 |
+
| ⚡ Best for | Low-end CPUs and tiny footprint | CPU ضعیف و کمترین اندازه |
|
| 46 |
+
|
| 47 |
- Canonical repository: [`Reza2kn/Shenava-Rizeh-Pizeh-v1.0`](https://huggingface.co/Reza2kn/Shenava-Rizeh-Pizeh-v1.0)
|
| 48 |
- PersianML mirror: [`PersianML/Shenava-Rizeh-Pizeh-v1.0`](https://huggingface.co/PersianML/Shenava-Rizeh-Pizeh-v1.0)
|
| 49 |
- Teacher: [`Reza2kn/Shenava-Rizeh-v1.0`](https://huggingface.co/Reza2kn/Shenava-Rizeh-v1.0)
|
| 50 |
|
| 51 |
+
## 🧠 Model contract
|
| 52 |
|
| 53 |
- Audio: mono, 16 kHz Persian speech.
|
| 54 |
- Encoder: `d_model=144`, 12 layers, 8x subsampling.
|
|
|
|
| 59 |
|
| 60 |
The release reported real-time FP32 tract inference on a 2015 Cortex-A7 (RTF about 0.91). Treat that as a release-specific device measurement, not a universal latency guarantee.
|
| 61 |
|
| 62 |
+
## 📊 Published evaluation
|
| 63 |
|
| 64 |
Decoded with context `[70,13]` and the double-benchmark ITN/Persian-digit normalization convention.
|
| 65 |
|
|
|
|
| 68 |
| visualears-golden-6669 | 24.55% | 8.89% |
|
| 69 |
| FLEURS-fa | 26.95% | 10.22% |
|
| 70 |
|
| 71 |
+
## 🚀 Load with NeMo
|
| 72 |
|
| 73 |
```python
|
| 74 |
from nemo.collections.asr.models import ASRModel
|
|
|
|
| 79 |
|
| 80 |
Choose this model when footprint and low-end CPU viability matter more than the accuracy available from the 32M Rizeh or 114M Koochik checkpoints.
|
| 81 |
|
| 82 |
+
## 🇮🇷 خلاصهٔ فارسی
|
| 83 |
|
| 84 |
«شنوا ریزهپیزه» کوچکترین مدل خانواده است: ۶٫۹ میلیون پارامتر برای اجرای کمهزینه روی CPUهای ضعیف. این مخزن checkpoint اصلی FP32 و NeMo را نگه میدارد؛ اندازهٔ کم با افت دقت نسبت به ریزه و کوچیک همراه است.
|
| 85 |
|
| 86 |
+
## 🌌 Explore Shenava-1
|
| 87 |
+
|
| 88 |
+
[🧠 Koochik 114M](https://huggingface.co/Reza2kn/Shenava-Koochik-v1.0) · [⚖️ Rizeh 32M](https://huggingface.co/Reza2kn/Shenava-Rizeh-v1.0) · **🐣 Rizeh-Pizeh 6.9M**
|
| 89 |
+
|
| 90 |
Apache-2.0. Accuracy varies with accent, noise, overlap, recording channel, and code-switching.
|