OpenVoiceOS/malaysian-whisper-medium-onnx

malaysian-whisper-medium — ONNX

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

License: the source model card does not specify a license explicitly; this is a straight format-conversion mirror of the public weights, with attribution to mesolitica preserved. If you are the author and want a specific license applied, please open an issue/PR.

First specialized ONNX ASR model for Malay 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="ms"))

Verified on a FLEURS Malay (ms_my) test clip:

• Reference: "beliau tidak dapat mengambil ubat yang diperlukan untuk menahan sakitnya kerana ia telah diharamkan dari perlawanan"
• fp32: "Beliau tidak dapat mengambil ubat yang diperlukan untuk menahan sakitnya kerana ia telah diharamkan dari pelawanan." (RTF 2.36)
• int8: same transcript (RTF 1.35)

Both fluent, matching the reference (one minor "per-" vs "pe-" difference in the last word). 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).