DeadBranches/acft-whisper-large-v3-turbo_q8_0

Model Card

Model Description

This is a whisper.cpp quantization of the FUTO acft whisper large-v3-turbo finetune performed by MahmoudAshraf. series of OpenAI's Whisper models.

FUTO acft models have been finetuned for dynamic audio context robustness, allowing shorter audio contexts for better performance with short audio inputs. The method is detailed in our GitHub repo.

• License: Apache-2.0
• Finetuned from model: OpenAI Whisper

Quantization Process

The original .safetensor model was converted to ggml format and then quantized using tools provided with the whisper.cpp repo.

>conda activate whisper.cpp

> WHISPERCPP_REPO=~/whisper.cpp
> WHISPER_PACKAGE=~/python-packages # Location of pip `whisper` package
> python "$WHISPERCPP_REPO/models/convert-h5-to-ggml.py" ~/hf-models/acft-whisper-large-v3-turbo "$WHISPER_PACKAGE" ~/quantized-models/acft-whisper-large-v3-turbo_q8_0

> cd ~/quantized-models/acft-whisper-large-v3-turbo_q8_0
> mv ggml-model.bin acft-whisper-large-v3-turbo-f16.bin

> python "$WHISPERCPP_REPO/bin/quantize.exe" acft-whisper-large-v3-turbo-f16.bin acft-whisper-large-v3-turbo-q8_0.bin q8_0
ggml_common_quantize_0: model size  =  3085.62 MB
ggml_common_quantize_0: quant size  =   833.08 MB | ftype = 7 (q8_0)
main: quantize time = 20616.65 ms
main:    total time = 20616.65 ms

Uses

These models are not useful by themselves under default Whisper runtime configurations.

The easiest way to test differing audio context is to use whisper.cpp with the --audio-context parameter. We provide converted whisper.cpp models in our GitHub README.

Other Information

More information can be found in this GitHub README.