Downloads · 30 days
0
mah92/sherpa-onnx-nemo-ctc-fa-shenava-koochik-v1.0-streaming-2026-06-26
sherpa-onnx-nemo-ctc-fa-shenava-koochik-v1.0-streaming-2026-06-26 is a automatic speech recognition model from mah92. Use it when you need speech turned into text. It is set up for sherpa-onnx. The card lists the license as cc-by-nc-4.0.
Koochik (کوچیک, "small") is the 114M teacher / flagship of the Shenava‑1 family — a FastConformer Hybrid RNNT/CTC model fine‑tuned on clean Persian data with vetokv4. This repository contains the CTC streaming FP32 ON…
Downloads · 30 days
0
Access
Public
Updated Jul 31, 2026
Repo size
459 MB
Likes
0
Public
Click a slice to open those files.
.onnx459 MB · 100%
From the Hugging Face model README
Koochik (کوچیک, "small") is the 114M teacher / flagship of the Shenava‑1 family — a FastConformer Hybrid RNNT/CTC model fine‑tuned on clean Persian data with ve_tok_v4. This repository contains the CTC streaming FP32 ONNX export for use with sherpa-onnx.
The original NeMo checkpoint is at Reza2kn/Shenava-Koochik-v1.0. This export uses the CTC decoder head only.
A knowledge‑distillation cascade of on‑device Persian ASR models — one teacher distilled down to a 6.9M student. This ONNX export is one member of the family; the parent models are:
| Member | Size | HF Repo |
|---|---|---|
| Koochik v1.0 (teacher) | 114M | Reza2kn/Shenava-Koochik-v1.0 |
| Rizeh v1.0 (mid student) | 32M | Reza2kn/Shenava-Rizeh-v1.0 |
| Rizeh Pizeh v1.0 (tiny student) | 6.9M | Reza2kn/Shenava-Rizeh-Pizeh-v1.0 |
ITN + Persian‑digit normalizer (the double‑benchmark convention), decoded at att_context_size=[70,13].
| Member | golden‑6669 WER | CER | FLEURS‑fa WER | CER |
|---|---|---|---|---|
| Koochik v1.0 (114M) | 7.49% | 2.30% | 10.64% | 3.79% |
| Rizeh v1.0 (32M) | 12.11% | 3.94% | 14.45% | 5.10% |
| Rizeh Pizeh v1.0 (6.9M) | 24.55% | 8.89% | 26.95% | 10.22% |
Koochik v1.0 is #2 on the public double‑benchmark, behind only cloud Gemini — the best on‑device Persian ASR, beating a 1.5B Whisper‑Persian by >2× WER at 1/13 the size.
Note: These benchmark numbers are from the original NeMo model. The ONNX CTC export may show slight degradation due to the removal of the transducer head and ONNX export optimizations.
| Property | Value |
|---|---|
| Parent Model | Reza2kn/Shenava-Koochik-v1.0 |
| Model Type | EncDecHybridRNNTCTCBPEModel (CTC head only) |
| Architecture | FastConformer Hybrid RNNT/CTC → CTC decoder export |
| Language | Persian (Farsi) |
| Parameters | 114M |
| Vocabulary Size | 1025 tokens (BPE + blank) |
| Tokenizer | ve_tok_v4 — SentencePiece BPE‑1024, digit/punct/«»‑aware. Numbers are spoken‑form; apply ITN at display for digits. |
| Sample Rate | 16 kHz |
| d_model | 512 / 17 layers |
| dw_striding | ×8 (80 ms/frame) |
| Subsampling | 8× |
| Multi‑latency | [[70,13],[70,6],[70,1],[70,0]] |
| Streaming | ✅ Yes (with cache support) |
| Weight Type | FP32 (full precision) |
| Author | Reza2kn |
| Version | 1.0 |
| Variant | Directory | Type | Precision |
|---|---|---|---|
| Streaming FP32 | sherpa-onnx-shenava-koochik-v1.0-ctc | Streaming | FP32 |
| Streaming INT8 | sherpa-onnx-shenava-koochik-v1.0-ctc-int8 | Streaming | INT8 quantized |
| Non-streaming FP32 | sherpa-onnx-shenava-koochik-v1.0-ctc-non-streaming | Non-streaming | FP32 |
| Non-streaming INT8 | sherpa-onnx-shenava-koochik-v1.0-ctc-non-streaming-int8 | Non-streaming | INT8 quantized |
.
├── README.md
├── model.onnx # ONNX model (FP32)
└── tokens.txt # Token vocabulary (1025 tokens)
import sherpa_onnx
recognizer = sherpa_onnx.OfflineRecognizer.from_sherpa_onnz(
tokens="tokens.txt",
encoder="model.onnx",
num_threads=4,
provider="cpu", # or "cuda", "coreml"
)
#include "sherpa-onnx/c-api/c-api.h"
SherpaOnnxOnlineRecognizerConfig config;
config.model_config.tokens = "tokens.txt";
config.model_config.encoder = "model.onnx";
config.model_config.num_threads = 4;
const SherpaOnnxOnlineRecognizer *recognizer =
SherpaOnnxOnlineRecognizerCreate(&config);
sherpa-onnx-offline \
--tokens=tokens.txt \
--encoder=model.onnx \
--num-threads=4 \
/path/to/test.wav
-int8 variant.Exported from a NeMo checkpoint.
CC-BY-NC 4.0 (inherited from the parent model)