rafdiraf/whisper-medium-id-ggml

Whisper Medium Indonesian — GGML Q5_0

This is a GGML Q5_0 quantized conversion of cahya/whisper-medium-id, an Indonesian-language fine-tune of OpenAI's Whisper Medium.

Details

Item	Value
Base model	cahya/whisper-medium-id (Apache 2.0)
Format	GGML (whisper.cpp compatible)
Quantization	Q5_0
File	ggml-model-medium-id-q5_0.bin
Size	539,212,467 bytes (~514 MB)
Original training data	Common Voice 11, Magic Data, TITML, Google FLEURS

Usage

This model is designed to be used with whisper.cpp or any application built on whisper.cpp bindings (e.g. whisper-rs).

Direct download URL (unauthenticated):

https://huggingface.co/rafdiraf/whisper-medium-id-ggml/resolve/main/ggml-model-medium-id-q5_0.bin

With whisper.cpp CLI

./main -m ggml-model-medium-id-q5_0.bin -l id -f audio.wav

With Notulenku

This model is available in Notulenku's model catalog as medium-id-q5_0. Select it from Settings → Transcription → Model.

Reported Accuracy (upstream)

From the original cahya/whisper-medium-id model card:

Dataset	WER
Common Voice 11 (test)	3.83%
Google FLEURS (test)	9.74%

Note: These benchmarks are on formal/read speech. Performance on spontaneous or informal Indonesian speech (e.g. meetings) will be significantly higher — see arXiv 2410.08828 for IDSV benchmark results.

Attribution & License

This model is a derivative of cahya/whisper-medium-id by Cahya Wirawan, licensed under the Apache License 2.0.

Changes made: Converted from Hugging Face Transformers format to GGML format using whisper.cpp's conversion scripts, then quantized to Q5_0.

Per Apache 2.0 §4 (Redistribution), this derivative work retains the original license and carries this attribution notice. No other modifications were made to the model weights beyond format conversion and quantization.