README: update for style-input + BOS/EOS wrap requirement + speed input + 25-voicepack catalog

ONNX: rename ref_s → style input for kokoro-js / kokoro-onnx interop (still v0.2 weights)

voices: bulk-add 25 deployed voicepacks (4 trained + 19 stock-Kokoro crossovers + 2 synthetic)

v0.2 ONNX with speed input — exposes speed as float32[1] dynamic input that scales pred_dur

revert: restore v0.2 TORCH26 ONNX as primary model.onnx (v0.5 has eat issues with mukta voicepack on WebGPU)

upgrade: v0.5 kokorobase ep4 (Stage-2-only-from-Kokoro recipe, FiLM baked into LSTM)

re-export ONNX with torch 2.6 (legacy tracer broke on torch >= 2.9)

Upload voices/mm_dnyanesh.pt with huggingface_hub

Upload voices/mf_mukta.pt with huggingface_hub

Upload onnx/model.onnx with huggingface_hub

re-extract Asha+Vivek from trim_silence-cleaned Rasa source

Stage 2 final (epoch_2nd_00009) — v0.1 release

voicepacks re-extracted from Stage 2 final

re-extract voicepacks from Stage 2 epoch 5 checkpoint

update fp32 ONNX from Stage 2 epoch 5 (epoch_2nd_00004)

rename q8 variant to model_quantized.onnx for transformers.js dtype='q8'

add int8 variant (283 MB, decoder Conv excluded — clean audio)

config: add model_type/architectures for transformers.js

v0.1-preview refresh: Stage 2 epoch 4 ONNX export + 4 voicepacks

v0.1-preview: ONNX export of Stage 2 epoch 2 checkpoint (fp32, 325 MB)

initial commit