OpenVoiceOS/whisper-small-bhojpuri-onnx

whisper-small-bhojpuri-merged — ONNX

ONNX export of Aditya109/whisper-small-bhojpuri-merged (Whisper small fine-tuned for Bhojpuri by Aditya109) for onnx-asr (standard whisper model type — works with stock onnx-asr, no patches needed). fp32 and int8 variants included.

License: cc-by-4.0, inherited from the source model.

First specialized ONNX ASR model for Bhojpuri in this collection.

Usage

Whisper has no dedicated bho language token; use language="hi" (Hindi, the closest supported Whisper token) — Bhojpuri behavior comes from the fine-tune itself. The tokenizer in this repo was not published separately; since vocab_size (51865) matches stock multilingual Whisper exactly, the standard openai/whisper-small tokenizer files are bundled here (vocab-size-verified).

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

Verified on a clip from ai4bharat/Rural_Women_Bhojpuri (FLEURS does not cover Bhojpuri):

• Reference: "किरपा करके हमरा डेबिट कार्ड से तिरानबे हजार रुपिया भेज दिही कार्ड नमर"
• fp32 (RTF 0.66): "खिरपा करके हमरा डेवेट debit कार्ड card से तिरान तिरान बे हजार रुपया भेजदे ही कार्ड card नंबर number"
• int8 (RTF 0.33): identical to fp32

Verified with caveats, honestly: both precisions produce the identical transcript, so quantization introduces no additional error here. However the transcript itself has real issues — repeated syllables on the number ("तिरान तिरान बे"), and duplicated English glosses appended after several Devanagari words ("कार्ड card", "नंबर number"). This looks like a genuine behavior of this checkpoint rather than an export artifact. Use with that caveat in mind. 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).