PMPR Kokoro bundle — 10s + 15s buckets
Unmodified subset of mattmireles/kokoro-coreml
(Apache-2.0) at revision 9b6c8db. 300 MB rather than 1,028 MB.
Why two buckets. Kokoro's decoders are duration-bucketed and a chunk is synthesized by the smallest bucket that fits, so bucket choice sets both speed and how long a chunk may be. Measured on the consuming app's real feedback: the 10 s bucket holds only 62% of its sentences whole, so the other 38% were being cut mid-sentence at a word boundary — an audible pause between two words. The 15 s bucket holds 100%. Shipping both keeps short sentences on the fast bucket and stops long ones being split.
Built with HF provenance verification; --allow-local-provenance produces a
bundle KokoroSDKModelProvider refuses to load.
- Downloads last month
- 4
Model tree for padejumo/pmpr-kokoro-models
Base model
yl4579/StyleTTS2-LJSpeech Finetuned
hexgrad/Kokoro-82M Quantized
mattmireles/kokoro-coreml