whisper-amharic-small-v2 — ONNX

ONNX export of chappM/whisper-amharic-small-v2 (Whisper small fine-tuned for Amharic by chappM) for onnx-asr (standard whisper model type — works with stock onnx-asr, no patches needed). fp32 and int8 variants included.

License: openrail, inherited from the source model.

First specialized ONNX ASR model for Amharic 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="am"))

Verified on a FLEURS Amharic (am_et) test clip (a sentence with two number sequences, "117" and "45"):

• Reference: "በሴቶች ቡድን ውስጥ አራት የበረዶ ተንሸራታቾች ሩጫቸውን መጨረስ አልቻሉም በግዙፉ ስላሎም ካሉት ከ117ቱ የበረዶ ተንሸራታቾች 45ቱ በውድድሩ ደረጃ ማግኘት አልቻሉም"
• fp32 (RTF 0.94): correct through the first clause, then degrades into a repetition loop around the digit-heavy portion of the sentence.
• int8 (RTF 0.30): also correct through the first clause, degrades less severely on the same digit-heavy portion, and terminates instead of looping.

Known limitation, not a quantization artifact: this checkpoint struggles with digit sequences spoken as words ("one hundred seventeen", "forty-five") in this test clip, in both precisions — likely a training-data gap in the base fine-tune rather than something introduced by this ONNX conversion. The first ~half of the sentence transcribes correctly and fluently in both variants; treat this export as verified-with-caveats rather than fully clean. 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).