---
library_name: onnx
base_model: Reza2kn/Shenava-Rizeh-0.9
base_model_relation: quantized
license: cc-by-4.0
tags:
- automatic-speech-recognition
- persian
- farsi
- fastconformer
- onnx
- fp16
- webgpu
---

# Reza2kn/Shenava-Rizeh-0.9-onnx-fp16

ONNX FP16 export of `Reza2kn/Shenava-Rizeh-0.9` for VisualEars browser/WebGPU inference.

## Runtime Contract

- Input `processed_signal`: `float16`, shape `[batch, 80, 2005]`
- Input `processed_signal_length`: `int64`, shape `[batch]`
- Output `logits`: `float16`
- Output `encoded_lengths`: `int64`
- Streaming attention context: `[70, 1]`
- Tokenizer: SentencePiece BPE, 1024 tokens plus CTC blank

## Validation

Parity smoke on the VisualEars 269 benchmark slice:

40/40 exact transcript matches, 1.000000 frame argmax agreement, 0.950 non-empty ONNX rate.

The comparison is source NeMo fp32 encoder plus auxiliary CTC head versus exported ONNX FP16 using the same preprocessed features.

## Files

- `shenava_rizeh_0_9_ctc_fixed2005_len_att70_1_fp16_full_io.onnx` — 6,171,937 bytes — sha256 `114ba96038eb45d92f449fcc5fdbf67e4c09aa96cbc5b4a4fddb77a52e09df9e`
- `shenava_rizeh_0_9_ctc_fixed2005_len_att70_1_fp16_full_io.onnx.data` — 52,731,906 bytes — sha256 `3261ed46ccef096a213e29c10f4f7a7a629c2a5895412003ada4db0c1fa9940f`
- `shenava_rizeh_0_9_ctc_fixed2005_len_att70_1_fp16_full_io_embedded.onnx` — 58,875,455 bytes — sha256 `7490c7d364f869912c57c1d57f8ee8688fa284af1305ff8d333ac69c65a74c25`
- `tokens.json` — 15,115 bytes — sha256 `895dc224c2b726fa988bdb9f4bd61a5334d594d9f37174399a2f97e68cad91ff`
- `preprocessor.json` — 1,795 bytes — sha256 `b08dcbb46aa54e4414c55306571530b05730b9df0472f7a13a6bcce19cdfaa61`
- `mel_filters_slaney_80x257.json` — 91,115 bytes — sha256 `327ad485dfcf1cbd9405ea6512aa0a788990a5c98ef14ba8585896cdc9749866`

Generated at `2026-06-23T02:04:27Z`.