Modelpublic

JoaoZaokk/kb-whisper-large-ggml

sourceHugging Faceapache-2.0updated 13d agoView on Hugging Face

0likes

Model Card

kb-whisper-large-ggml

Quantizations of the authors' own whisper.cpp f16 conversion of KBLab/kb-whisper-large, in q40 / q50 / q8_0 (phones use q4/q5, Macs q8).

Source checkpoint: KBLab/kb-whisper-large by KBLab · License: apache-2.0 (unchanged; this repo only re-packages the weights) Engine: Load with whisper.cpp (whisper-cli -m <file>), or any app that embeds it.

Files

File	Quantization	Size
`ggml-kb-whisper-large-q8_0.bin`	q8_0	1657 MB
`ggml-kb-whisper-large-q5_0.bin`	q5_0	1081 MB
`ggml-kb-whisper-large-q4_0.bin`	q4_0	889 MB

f16 is the lossless conversion; q8_0 is nearly identical in accuracy at ~55 % of the size; q5_0/q5_k are the phone-friendly choice; q4_* is smallest with a small accuracy cost.

How these were made

Converted from the upstream checkpoint with the engine's own converter, then quantized with the engine's quantizer. Each variant was checked by transcribing short Portuguese and English samples before upload.

Attribution

Weights are derivative works of the upstream model and keep its license. Please cite the original authors (KBLab). The conversion and hosting here are maintained by JoaoZaokk so that the download links used by the Odysseus / Open WebUI native apps stay stable. No warranty.