license: mit
quantized_by: Pomni
language:
- en
base_model:
- distil-whisper/distil-large-v3.5
pipeline_tag: automatic-speech-recognition
datasets:
- mozilla-foundation/common_voice_17_0
- openslr/librispeech_asr
- facebook/voxpopuli
- LIUM/tedlium
- MLCommons/peoples_speech
- speechcolab/gigaspeech
- edinburghcstr/ami
- espnet/yodas
tags:
- whisper.cpp
- ggml
- whisper
- audio
- speech
- voice
- distil
Distil-Large-v3.5 quants
This is a repository of GGML quants for distil-large-v3.5 (a Whisper-based transcription model), for use with whisper.cpp.
If you are looking for a program to run this model with, then I would recommend EasyWhisper UI, as it is user-friendly, has a GUI, and automates a lot of the hard stuff for you.
List of Quants
Clicking on a link will download the corresponding quant instantly.
| Link | Quant | Size | Notes |
|---|---|---|---|
| GGML | F32 | 3.03 GB | Likely overkill. |
| GGML | F16 | 1.52 GB | Performs better than Q8_0 for noisy audio and music. |
| GGML | Q8_0 | 818 MB | Sweet spot; superficial quality loss at nearly double the speed. |
| GGML | Q6_K | 637 MB | |
| GGML | Q5_K | 538 MB | |
| GGML | Q5_1 | 585 MB | |
| GGML | Q5_0 | 538 MB | Last "good" quant; anything below loses quality rapidly. |
| GGML | Q4_K | 444 MB | Might not have lost too much quality, but I'm not sure. |
| GGML | Q4_1 | 491 MB | |
| GGML | Q4_0 | 444 MB | |
| GGML | Q3_K | 345 MB | |
| GGML | Q2_K | 269 MB | Completely non-sensical output. |
The F16 quant was taken from distil-whisper/distil-large-v3.5-ggml/ggml-model.bin.
Questions you may have
Why do the "K-quants" not work for me?
My guess is that your GPU might be too old to recognize them, considering that I have gotten the same error on my GTX 1080. If you would like to run them regardless, you can try switching to CPU inference.
Are the K-quants "S", "M", or "L"?
The quantizer I was using was not specific about this, so I do not know about this either.
What program did you use to make these quants?
I used whisper.cpp v1.7.6 on Windows x64, leveraging CUDA 12.4.0. For the F32 quant, I converted the original Hugging Face (H5) format model to a GGML using the models/convert-h5-to-ggml.py script.
One or multiple of the quants are not working for me.
Open a new discussion in the community tab about this, and I will look into the issue.