Skip to content

Provenance: ve.safetensors

The utterance voice encoder that drives voice cloning. This page records what it is, where it came from, and the terms it is redistributed under.

fieldvalue
fileve.safetensors
size5,695,784 bytes
sha256f0921cab452fa278bc25cd23ffd59d36f816d7dc5181dd1bef9751a7fb61f63c
contents16 tensors, float32: a 3-layer LSTM (lstm.weight_ih_l{0,1,2}, lstm.weight_hh_l{0,1,2}, biases) and a linear projection
what it does40-mel partials of a ≤10 s clip → a 256-d utterance embedding
used byloudkit.enroll only. Synthesis does not touch it; a checkpoint without it speaks every voice in the library and cannot make a new one.

It is the VoiceEncoder of the Chatterbox T3 architecture, published by Resemble AI under the MIT licence:

The weights were not retrained. They were re-exported: read out of the upstream artefact and written to safetensors by this project’s own export script.

The same export path produced the other enrollment components named in NOTICE: the speech tokenizer, the CAM++ speaker encoder, and the T3 conditioning layers. NOTICE lists their upstream licences.

MIT, the same terms as the upstream weights.

That is the conservative reading. MIT permits sublicensing, so a derivative may in principle be distributed under Apache-2.0, and this project’s code is Apache-2.0 for the express patent grant. Relicensing somebody else’s MIT weights means asserting terms over an artefact whose substance is theirs, and it buys only tidiness in a table. The weights stay MIT, the code stays Apache-2.0, and README.md says so where a redistributor reads it before NOTICE.

The full MIT text and Resemble AI’s copyright line are in NOTICE, which ships inside the pip, npm, crates.io and Go packages rather than only in this repository.

The upstream revision. The export ran against a local checkout of the Chatterbox weights. The revision is not recorded in the artefact metadata or in the export script. The hash above identifies this file. It does not identify which upstream snapshot produced it.

The encoder is a read-out rather than a transform, so a re-export from a published Chatterbox revision is expected to produce a byte-identical file. That is the route for anyone who needs the link established.