Quantized Hebrew models
Pre-release
Pre-release
·
5 commits
to claude/android-meeting-transcriber-app-ffvbe7
since this release
Quantized builds of the ivrit.ai Hebrew Whisper models, for whisper.cpp.
Source: https://huggingface.co/ivrit-ai/whisper-large-v3-turbo-ggml/resolve/main/ggml-model.bin (1549 MB, f16)
ivrit-whisper-large-v3-turbo-q5_0.bin— 547 MB, sha2566c1da92e8e41dd64b8cc402eee7eb7a433d2152567e1a4d9cf181fefcc67a572
q8_0 is near-lossless and roughly halves the memory traffic;
q5_0 is about a third of the original at some cost in accuracy.