kale-tts-1-tiny / pytorch /README.md
NERDDISCO's picture
Clear file names: kale-tts-1-tiny.onnx at the root, PyTorch weights in pytorch/
6901100 verified
|
Raw History Blame Contribute Delete
1.24 kB
<!-- SPDX-License-Identifier: Apache-2.0 | Copyright 2026 Kale -->
# kale-tts-1-tiny training checkpoints (PyTorch)
The trained parts `kale-tts-1-tiny.onnx` was exported from, for fine-tuning with `recipe/`. Weights only (no optimiser states), Apache-2.0.
The teacher both models were distilled from is `teacher/kale-tts-1-teacher.pth` (Kokoro-82M fine-tuned on the 10 voices), with its
per-voice style packs in `teacher/voices/<id>.pt`; tiny keeps its timing-and-melody planner, frozen.
| file | what |
|---|---|
| `tiny/ettin_text_side_fle.pt` | tiny's text side: a 3.0M phoneme reader next to the frozen Ettin encoder (jhu-clsp/ettin-encoder-17m), distilled from kale-tts-1-base |
| `tiny/pitch_heads_flp.pt` | pitch / loudness heads + per-voice offsets `pv` (10 x 128), trained on the voices' Maya-1 audio |
| `tiny/duration_head_fld.pt` | duration head (pace windows), `pv` frozen from the pitch heads |
| `tiny/decoder_flkd.pt` | tiny's 3.65M sound generator, 10 style rows, with its two critics (`model`, `D`, `J`) |
Voice order in every per-voice table: s057b, f06, d03, w165, w602, m3, t071, d13, x328, h013f.
`tiny/earlier/` holds the earlier training run's copies of the same parts (the `t10*` files), kept for reference.