Shenava — Rizeh Pizeh v1.0 (6.9M) · tract cache‑aware streaming (fp32)
fp32 cache‑aware streaming ONNX of Shenava-Rizeh-Pizeh-v1.0 — real‑time on a 2015 Cortex‑A7 (RTF ≈ 0.91) via the tract Rust engine. Per‑step: processed_signal [1,80,32] + caches → ctc_logits [1,3,1025] + next caches (chunk 4, shift 3, left 64).
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 recordReza2kn/Shenava-Rizeh-v1.0— Rizeh v1.0 (32M) · mid-tier studentReza2kn/Shenava-Rizeh-Pizeh-v1.0— Rizeh Pizeh v1.0 (6.9M) · tiniest — real-time on a 2015 Cortex-A7 ◀ this model (or its parent)
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.
fp32, not a quant — FastConformer sub‑fp16 in tract was never usable; the A7 real‑time result is fp32.
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.
Model tree for Reza2kn/Shenava-Rizeh-Pizeh-v1.0-tract
Base model
nvidia/stt_fa_fastconformer_hybrid_large