---
language: [fa]
license: apache-2.0
library_name: onnx
tags: [automatic-speech-recognition, fastconformer, persian, fa, onnx, fp16, on-device, webgpu, ctc]
base_model: Reza2kn/shenava-fa-fastconformer-115m
---
# Shenava v4 — FastConformer-115M CTC — fp16 ONNX (in-browser / on-device)

fp16 ONNX export of the **CTC head** of [Reza2kn/shenava-fa-fastconformer-115m](https://hf.co/Reza2kn/shenava-fa-fastconformer-115m), for `onnxruntime-web` (WebGPU / WASM) and `onnxruntime`. Powers the offline Persian captioner at **VisualEars.com**.

- **Input:** `processed_signal` float32 `[1, 80, 2005]` — 80-dim log-mel features over a fixed 20.05 s window (NeMo FastConformer preprocessor; configs `115m_preprocessor.json` + `115m_mel_filters.json` included).
- **Outputs:** `logits` `[1, 251, 1025]` (CTC, blank id 1024) and `encoded_lengths`. Greedy-CTC decode with `115m_tokens.json`; truncate logits to `ceil(real_mel_frames / 8)` before collapse.
- **Golden6669 WER:** **7.92%** (CTC, official Persian normalizer). The parent's RNNT head scores 7.29%.
- **Streaming:** the deployed app streams by full-context re-decode of a growing buffer — zero accuracy penalty.

Fully offline, on-device. Part of the VisualEars project (SLT 2026).