Polyvox per-language Whisper GGML mirror
Redundancy mirror of the per-language whisper.cpp models Polyvox can download. Every file is a q8_0 quantization of an upstream fine-tune. The weights are not ours and we changed nothing beyond format conversion and quantization.
Each artifact ships two files beside it: <id>.LICENSE.txt is the upstream
license text, and <id>.NOTICE.txt names the upstream model, its author, and
which conversion steps were applied.
Licenses differ per file, so the repo-level field is other. Read the
per-file license, not the repo field.
| File | Language | Upstream | License |
|---|---|---|---|
ggml-ivrit-he-turbo-q8_0.bin |
he | ivrit-ai/whisper-large-v3-turbo | Apache-2.0 |
ggml-kb-whisper-small-q8_0.bin |
sv | KBLab/kb-whisper-small | Apache-2.0 |
ggml-kb-whisper-medium-q8_0.bin |
sv | KBLab/kb-whisper-medium | Apache-2.0 |
ggml-kb-whisper-large-q8_0.bin |
sv | KBLab/kb-whisper-large | Apache-2.0 |
ggml-finnish-nlp-fi-large-v3-q8_0.bin |
fi | Finnish-NLP/whisper-large-finnish-v3 | Apache-2.0 |
ggml-adriszmar-es-turbo-q8_0.bin |
es | adriszmar/whisper-large-v3-turbo-es | MIT |
ggml-coriollon-ru-turbo-q8_0.bin |
ru | coriollon/whisper-large-v3-turbo-russian | Apache-2.0 |
ggml-turkspeak-tr-large-v3-q8_0.bin |
tr | RsGoksel/turkspeak-lv3-ITU_Mainframe | Apache-2.0 |
ggml-oddadmix-ar-turbo-q8_0.bin |
ar | oddadmix/whisper-large-v3-turbo-arabic-dialectal | Apache-2.0 |
ggml-phowhisper-vi-large-v2-q8_0.bin |
vi | vinai/PhoWhisper-large | BSD-3-Clause |
ggml-pathumma-th-large-v3-q8_0.bin |
th | nectec/Pathumma-whisper-th-large-v3 | Apache-2.0 |
ggml-localai-it-turbo-q8_0.bin |
it | LocalAI-io/whisper-large-v3-turbo-it | MIT |
Why q8_0
q8_0 is indistinguishable from f16 on these fine-tuned decoders while halving the download and the resident footprint. Lower quantizations lose real accuracy on non-English audio.