Update README: q6q8 replaces broken q4q6, add WER validation table

Publish quantized turbo (encoder 6.0-bit, decoder 8.5-bit) -- WER-validated on 30 FLEURS clips, matches fp16 baseline quality (de 3.4/en 15.5/fr 8.8/ja 2.6/ko 22.5/zh 9.0 vs fp16 2.8/15.5/7.8/2.6/18.1/9.0)

Remove broken quantized build -- on-device WER test showed severe quality collapse (67-118% WER vs fp16 baseline). Superseded by dictionlabs_whisper-large-v3-turbo-q6q8.

Remove em dash

Simplify README for end users -- drop internal PSNR/CLI-flag build details

Rewrite README: drop Argmax mirroring language, list our own turbo fp16 + quantized q4q6 builds

Remove Argmax mirror -- no prod installs depend on DictionLabs/whisperkit-coreml yet, and we do not mirror Argmax content anymore

Remove intermediate no-suffix folder, superseded by dictionlabs_whisper-large-v3-turbo-q4q6

Remove stale Argmax-style-named folder, superseded by dictionlabs_whisper-large-v3-turbo-q4q6

Publish quantized turbo under final name dictionlabs_whisper-large-v3-turbo-q4q6

Rename quantized turbo folder to dictionlabs_whisper-large-v3-turbo

whisperkittools- generated files: openai_whisper-large-v3-turbo_546MB

docs: rewrite as Diction-owned build, drop mirror/1:1 framing

whisperkittools- generated files: openai_whisper-large-v3-turbo

Card: per-variant provenance; turbo is redistributed from Argmax, not ours

Mirror openai_whisper-large-v3-v20240930_turbo_632MB from argmaxinc/whisperkit-coreml@97a5bf9bbc74

Add model card: usage, variants, provenance

whisperkittools- generated files: openai_whisper-small

Upload folder using huggingface_hub

initial commit