Reza2kn/Shenava-Koochik-v1.0

Shenava — Koochik v1.0 (114M) · Persian streaming ASR

Koochik (کوچیک, “small”) is the 114M teacher / flagship of the Shenava‑1 family — a FastConformer Hybrid RNNT/CTC model (CTC head deployed) fine‑tuned on clean Persian data with ve_tok_v4. It powers shenava.app and is the teacher the 32M and 6.9M students distil from. This repo holds the fp32 NeMo source; quantized formats live in their own repos (below).

The Shenava‑1 family

A knowledge‑distillation cascade of on‑device Persian ASR models — one teacher distilled down to a 6.9M student. This model is one member; its siblings:

• Reza2kn/Shenava-Koochik-v1.0 — Koochik v1.0 (114M) · teacher / flagship — on-device WER record ◀ this model (or its parent)
• Reza2kn/Shenava-Rizeh-v1.0 — Rizeh v1.0 (32M) · mid-tier student
• Reza2kn/Shenava-Rizeh-Pizeh-v1.0 — Rizeh Pizeh v1.0 (6.9M) · tiniest — real-time on a 2015 Cortex-A7

Benchmark — fair WER/CER

ITN + Persian‑digit normalizer (the double‑benchmark convention), decoded @ 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.

Quantized formats (own repos, children of this model)

• Shenava-Koochik-v1.0-ONNX-fp16 — ONNX fp16 (the shenava.app deploy format)
• Shenava-Koochik-v1.0-CoreML-fp16 — CoreML fp16 mlprogram

114M, d_model 512 / 17 layers, dw_striding ×8 (80 ms/frame), multi‑latency [[70,13],[70,6],[70,1],[70,0]].

Tokenizer: ve_tok_v4 (SentencePiece BPE‑1024 +blank, digit/punct/«»‑aware). Numbers are spoken‑form; apply ITN at display for digits. Part of VisualEars / Shenava.