Reza2kn/Shenava-Rizeh-Pizeh-v1.0-sherpa-onnx

Shenava Rizeh-Pizeh v1.0 — Persian ASR for sherpa-onnx (6.9M)

Offline Persian ASR (NVIDIA FastConformer, CTC head) exported for sherpa-onnx. Runs on every sherpa-onnx target: Python, C++, Android, iOS, WebAssembly, C#, Go, Swift, Kotlin, Rust.

Files: model.onnx + tokens.txt (16 kHz). Quality in sherpa-onnx: runs (tiny model, ~24.5% WER).

Usage (Python — identical API on all platforms)

import sherpa_onnx, soundfile as sf, numpy as np
from persian_itn import itn   # spoken-number -> Persian-digit post-processor

rec = sherpa_onnx.OfflineRecognizer.from_nemo_ctc(
    model="model.onnx", tokens="tokens.txt", num_threads=4)
y, sr = sf.read("audio.wav"); y = np.asarray(y, np.float32)
s = rec.create_stream(); s.accept_waveform(sr, y); rec.decode_stream(s)
print(itn(s.result.text))     # ITN: هشت -> ۸

Notes

• This model consumes un-normalized log-mel and must not be per-feature normalized. The ONNX metadata sets normalize_type empty so sherpa-onnx skips its NeMo normalization — essential; with the default per-feature normalization the output is garbage.
• Numbers are spelled by design; apply the bundled persian_itn.py to the output to get digits.