Missing quantizations ggml-large-v3-q8_0.bin ggml-large-v1-q8_0.bin ggml-large-v1-q5_0.bin
#32
by RaivisDejus - opened
This repository is very valuable, thanks for it.
Notices come missing model quantizations that some might expect for consistency
Missing files:
- ggml-large-v3-q8_0.bin
- ggml-large-v1-q8_0.bin
- ggml-large-v1-q5_0.bin
You can generate these locally rather than waiting for them. whisper.cpp ships a quantize tool, so the f16 ggml model plus one command gives you the q8_0 or q5_0 variant.
Worth doing regardless, because it lets you check the accuracy trade-off on your own audio rather than trusting a general benchmark. Quantisation tends to cost more on accented speech and noisy recordings than a clean test set suggests.