🎙️ Shenava Koochik v1.0 · شنوا کوچیک
The 114M-parameter flagship of the Shenava-1 Persian ASR family. This is the FP32 NeMo source checkpoint for evaluation, fine-tuning, and export; use one of the deployment repositories below for browser, Core ML, sherpa-onnx, or tract applications.
✨ At a glance | معرفی سریع
| English | فارسی | |
|---|---|---|
| 🧠 Role | 114M flagship teacher | مدل معلم و پرچم‌دار ۱۱۴M |
| 📦 Format | FP32 NeMo source | checkpoint اصلی FP32 و NeMo |
| 🎧 Input | 16 kHz mono Persian speech | گفتار فارسی تک‌کانالهٔ ۱۶ کیلوهرتز |
| 📝 Output | Persian transcription | رونویسی فارسی |
| 🛠️ Best for | Evaluation, fine-tuning, and export | ارزیابی، آموزش تکمیلی و تبدیل |
- Canonical repository:
Reza2kn/Shenava-Koochik-v1.0 - PersianML mirror:
PersianML/Shenava-Koochik-v1.0
🧠 Model
- Architecture: FastConformer Hybrid RNNT/CTC; the CTC head is used by the published deployment exports.
- Audio: mono, 16 kHz Persian speech.
- Encoder:
d_model=512, 17 layers, 8x subsampling (about 80 ms per encoder step). - Contexts:
[70,13],[70,6],[70,1], and[70,0]. - Tokenizer: ve_tok_v4, SentencePiece BPE-1024 plus CTC blank.
- Output: Persian text. Numbers are emitted in spoken form; apply Persian inverse text normalization when digits are wanted.
The two .nemo filenames are compatibility aliases with identical content and SHA-256:
f7b5124a9fd2d50c15bf070abdc0f80ec2449d64948c5657adb79a7e91dd1d16
📊 Published evaluation
Decoded with context [70,13] and the double-benchmark ITN/Persian-digit normalization convention.
| Set | WER | CER |
|---|---|---|
| visualears-golden-6669 | 7.49% | 2.30% |
| FLEURS-fa | 10.64% | 3.79% |
🚀 Load with NeMo
from nemo.collections.asr.models import ASRModel
model = ASRModel.restore_from("shenava-koochik-v1.0.nemo")
text = model.transcribe(["speech.wav"])[0].text
print(text)
🧩 Deployment variants
- ONNX FP16: fixed-window browser/WebGPU/WASM export used by Shenava.
- sherpa-onnx: offline cross-platform bundle.
- Core ML FP16: fixed-window ML Program package.
- Core ML iOS 15: cache-aware streaming NeuralNetwork model for older Apple OS versions.
- tract streaming and tract offline: pure-Rust runtime exports.
🇮🇷 خلاصهٔ فارسی
این مخزن نسخهٔ اصلی FP32 و NeMo مدل ۱۱۴ میلیون‌پارامتری «شنوا کوچیک» است. برای اجرای مرورگر، Core ML، sherpa-onnx یا tract از مخزن مخصوص همان قالب استفاده کنید. ورودی گفتار فارسی تک‌کانالهٔ ۱۶ کیلوهرتز و خروجی متن فارسی است.
🌌 Explore Shenava-1
🧠 Koochik 114M · ⚖️ Rizeh 32M · 🐣 Rizeh-Pizeh 6.9M · 🌐 Browser · 🍎 Apple · 🦀 Rust
Apache-2.0. ASR quality can vary with accent, noise, overlap, recording channel, and code-switching; review meaning-critical transcripts before relying on them.