language:
- fa
license: apache-2.0
library_name: nemo
pipeline_tag: automatic-speech-recognition
base_model:
- Reza2kn/Shenava-Rizeh-v1.0
base_model_relation: finetune
tags:
- automatic-speech-recognition
- speech
- persian
- farsi
- fastconformer
- ctc
- streaming
- on-device
- shenava
- shenava-1
- visualears
- nemo
- distillation
metrics:
- wer
- cer
datasets:
- Reza2kn/visualears-persian-asr-16k
- Reza2kn/visualears-golden-6669
- Reza2kn/fleurs-fa-benchmark
🐣🎙️ Shenava Rizeh-Pizeh v1.0 · شنوا ریزهپیزه
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.
✨ At a glance | معرفی سریع
| English | فارسی | |
|---|---|---|
| 🐣 Role | Smallest Shenava-1 model | کوچکترین مدل خانوادهٔ Shenava-1 |
| 🪶 Scale | 6.9M parameters | ۶٫۹ میلیون پارامتر |
| 📦 Format | FP32 NeMo source | checkpoint اصلی FP32 و NeMo |
| 🧠 Lineage | Koochik → Rizeh → Rizeh-Pizeh | زنجیرهٔ تقطیر کوچیک ← ریزه ← ریزهپیزه |
| ⚡ Best for | Low-end CPUs and tiny footprint | CPU ضعیف و کمترین اندازه |
- Canonical repository:
Reza2kn/Shenava-Rizeh-Pizeh-v1.0 - PersianML mirror:
PersianML/Shenava-Rizeh-Pizeh-v1.0 - Teacher:
Reza2kn/Shenava-Rizeh-v1.0
🧠 Model contract
- Audio: mono, 16 kHz Persian speech.
- Encoder:
d_model=144, 12 layers, 8x subsampling. - Contexts:
[70,13],[70,6],[70,1], and[70,0]. - Deployed head: CTC.
- Tokenizer: ve_tok_v4, SentencePiece BPE-1024 plus blank.
- Output: Persian text; use display-layer ITN when Persian digits are required.
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.
📊 Published evaluation
Decoded with context [70,13] and the double-benchmark ITN/Persian-digit normalization convention.
| Set | WER | CER |
|---|---|---|
| visualears-golden-6669 | 24.55% | 8.89% |
| FLEURS-fa | 26.95% | 10.22% |
🚀 Load with NeMo
from nemo.collections.asr.models import ASRModel
model = ASRModel.restore_from("shenava-rizeh-pizeh-v1.0.nemo")
print(model.transcribe(["speech.wav"])[0].text)
Choose this model when footprint and low-end CPU viability matter more than the accuracy available from the 32M Rizeh or 114M Koochik checkpoints.
🇮🇷 خلاصهٔ فارسی
«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.
🌌 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.