NOTICE

# Third-party notices

ColorSplitter is MIT licensed (see `LICENSE`). It contains code and model
definitions derived from the following projects. Attribution is kept here and
repeated in the file headers where the code lives.

---

## Resemblyzer — MIT

<https://github.com/resemble-ai/Resemblyzer>

Used for:

* `src/colorsplitter/core/audio.py` — the `preprocess_wav`, `normalize_volume`,
  `trim_long_silences` and `wav_to_mel_spectrogram` implementations are inlined
  from `resemblyzer/audio.py`. The librosa calls are kept argument-for-argument
  identical so that embeddings stay compatible with existing checkpoints.
* `src/colorsplitter/core/hparams.py` — audio and model hyper-parameters.
* `src/colorsplitter/models/voice_encoder.py` — the `VoiceEncoder` network
  definition, including `compute_partial_slices` and `embed_utterance`.
* The optional `speaker-upstream-v1` weight (`resemblyzer/pretrained.pt`).

The `Resemblyzer` *package* is deliberately not a dependency: only the parts
actually used are vendored, so the install does not pull in the whole project
and its transitive requirements.

---

## 3D-Speaker — Apache License 2.0

<https://github.com/alibaba-damo-academy/3D-Speaker>

Used for:

* `src/colorsplitter/core/cluster.py` — `SpectralCluster` and `UmapHdbscan`,
  themselves adapted from SpeechBrain's spectral clustering implementation.

The clustering behaviour is preserved. `p_pruning` is vectorised and the probe
matrix is optionally sparsified for the eigensolver; `tests/test_cluster_equivalence.py`
pins the results against a verbatim copy of the original implementation.

---

## wav2vec2-large-robust-12-ft-emotion-msp-dim

<https://huggingface.co/audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim>

Used for the `emotion` encoder. The model weights are downloaded on demand and
are not redistributed here; see the model card for their terms.

`src/colorsplitter/models/wav2vec2.py` reimplements the wav2vec2 trunk against
that checkpoint. The graph was written with reference to HuggingFace
Transformers' `modeling_wav2vec2.py` and
`feature_extraction_wav2vec2.py` (Apache License 2.0,
<https://github.com/huggingface/transformers>), in particular the
weight-normalisation convention and the input normalisation, so that the
numerics match. Transformers is **not** a runtime dependency.

`src/colorsplitter/models/safetensors.py` is an independent minimal reader for
the safetensors format (see <https://github.com/huggingface/safetensors>,
Apache License 2.0), written to avoid pulling in another package.

---

## Encoder checkpoints

The bundled training lineage descends from CorentinJ's Real-Time-Voice-Cloning
(<https://github.com/CorentinJ/Real-Time-Voice-Cloning>, MIT), specifically its
speaker-encoder training and GE2E objective.