OpenVoiceOS/whisper-small-ps-augmented-onnx

whisper_small_ps_augmented — ONNX

ONNX export of ihanif/whisper_small_ps_augmented (Whisper small fine-tuned for Pashto by ihanif) for onnx-asr (standard whisper model type — works with stock onnx-asr, no patches needed). fp32 and int8 variants included.

License: apache-2.0, inherited from the source model.

First specialized ONNX ASR model for Pashto in this collection.

Usage

import onnx_asr
model = onnx_asr.load_model("whisper", "path/to/this/repo")  # or quantization="int8"
print(model.recognize("audio_16khz.wav", language="ps"))

Verified on a FLEURS Pashto (ps_af) test clip:

• Reference: "د نارینه په ښکته کولو کې ثابت اوسئ، او د خپلې ځمکې د ودریدو څخه مه ویریږئ کلتوري توپیرونه که نه، دا سم نه کوي!"
• fp32 (RTF 0.21): "نانینا په خطا کولو کثابت اوسئ د خپلې ځمکې د ودیدرودو هغه معیارېږئ کلتوري توپیرو نه تاسم نه کوي"
• int8 (RTF 0.09): "نا ننه په ښتا کولو کې ثابت اوسې د خپلې ځمکې د ویاریږی کلتوري توپیرونو نه تاسم نه کوي"

Verified with caveats, honestly: both variants get the general shape and several words right ("خپلې ځمکې", "کلتوري توپیرو", "ثابت اوسئ") but diverge noticeably from the reference on other words. fp32 and int8 are similarly rough — this looks like a genuine accuracy ceiling of this small fine-tuned checkpoint on this clip, not something introduced by the ONNX conversion or quantization (both precisions degrade about equally). Treat this as a correctly-converted mirror of the source model's real-world accuracy, not a claim of production-quality Pashto ASR. RTF measured on an AMD Ryzen 5 7600 (CPU, 4 OMP threads, shared/loaded box — not a clean benchmark number).

Int8 decoder was produced by quantizing the pre-merge decoders separately and re-merging (merge_decoders(..., strict=False)); direct quantization of the merged decoder graph does not shrink it (its If subgraphs are skipped by onnxruntime's dynamic quantizer).