emillykkegrann/hviske-v5-tiny-GGUF

hviske-v5-tiny GGUF (mirror)

This is an unmodified mirror of all 5 GGUF conversions of syvai/hviske-v5-tiny, created by syvai. All credit for the model belongs to them.

The original repository is gated, which means applications cannot download from it on a user's behalf. This mirror exists so Codictate can offer hviske-v5-tiny as a Danish speech model. Nothing about the weights has been changed.

Licence

Released under CC BY-NC 4.0, the same licence as the original. Non-commercial use only, attribution required. Codictate is free and open source software with no paid version and no commercial use.

• Original model: syvai/hviske-v5-tiny
• Original author: syvai
• Licence: CC BY-NC 4.0

Files

File	Size	Notes
hviske-v5-tiny-f16.gguf	503 MB	full precision, largest and slowest
hviske-v5-tiny-q8_0.gguf	268 MB	about half the size of f16
hviske-v5-tiny-q6_k.gguf	232 MB	smaller than q8_0
hviske-v5-tiny-q5_0.gguf	181 MB	smaller than q6_k
hviske-v5-tiny-q4_k.gguf	153 MB	smallest and fastest

The model card of the original repository reports the same Danish WER (10.51) for every one of these, so the choice is a size and speed trade-off rather than an accuracy one.

Codictate does not default to any one of these. All five are offered side by side, and the user picks the size and speed trade-off they want. If you want a starting point, hviske-v5-tiny-f16.gguf is full precision and the source repo's own primary, so it is the safest choice if you would rather not rely on the identical-WER claim above holding at the smaller Quantizations.

Runtime

These GGUFs load under CrispASR's cohere backend (--backend cohere).