Table of contents

Upstream source study inventory

Kokoro-Hexagon 0a03be39Updated 2026-09-26

These are complete top-level source trees under a local upstream mirror, checked out at the immutable commits below with shallow history. They are read-only study material. No source or output from these trees was added to the APK, model assembly, build inputs, or production runtime.

Tree Upstream Commit Study role
pytorch https://github.com/pytorch/pytorch.git 2b3ec34829036a65cd9d1398ea72a0167dc37470 Torch 2.14.0 operator and module semantics; matches the version named in lib/manifest.json
transformers https://github.com/huggingface/transformers.git 8ac2b916b042b1f78b75c9eb941c0f5d2cdd8e10 ALBERT and English G2P BART implementation; v4.57.0 study snapshot
numpy https://github.com/numpy/numpy.git a98529dfd92d73fd4d40ebaaf285cf2463bb92b1 Numerical behavior used by stock STFT construction
spacy https://github.com/explosion/spaCy.git 26b4d1dc04a812f426e4bef3e8a1b6f159d6f048 Misaki English tokenization and linguistic pipeline
num2words https://github.com/savoirfairelinux/num2words.git 07814cb114157f582c40a00119c2e9faba8dcee2 Misaki English number expansion
regex https://github.com/mrabarnett/mrab-regex.git 85e568c0e0be620b4ebe50c98ea92415f5995ffc Misaki Unicode token splitting
phonemizer https://github.com/bootphon/phonemizer.git ab8e780b302adb4e18f06350db3bc7f4564182c8 Misaki eSpeak fallback wrapper
espeak-ng https://github.com/espeak-ng/espeak-ng.git ba90c8e9f440ad544f674a790bb5f53878b6ffc5 Pronunciation engine source and data format
espeakng-loader https://github.com/thewh1teagle/espeakng-loader.git 0ddc87adf77e5850d7eeb542ac8a87d421b64daa Misaki's eSpeak library/data discovery
spacy-curated-transformers https://github.com/explosion/spacy-curated-transformers.git fe55b96afa1ec86a245c92095cf6559478b159a1 Misaki English spaCy pipeline configuration
aosp-frameworks-av https://android.googlesource.com/platform/frameworks/av e2f098935447ca4945946de5cb69db843fe3f003 (checkout); 9e7dd63dfff0cc967f025ea9e27a299aaa99fd69 (pinned study commit) Android AAudio implementation and PCM sink contract

The already mirrored canonical model sources remain kokoro at dfb907a02bba8152ca444717ca5d78747ccb4bec and misaki at fba1236595f2d2bf21d414ba6e57d25256afada3. The new snapshots were selected for source study; except for the Torch version, they are not yet asserted to match a version-locked numerical oracle environment. Before a parity claim depends on a framework implementation detail, pin the exact oracle package version and trace that version's source. Historical StyleTTS2, ONNX, QNN, and LLVM sources are not substitutes for the pinned Kokoro model.

The source chain is visible in kokoro/kokoro/model.py (ALBERT and model assembly), kokoro/kokoro/modules.py (CustomAlbert derives from upstream AlbertModel), kokoro/kokoro/istftnet.py (decoder), and misaki/misaki/en.py (English text path). Their original repositories are the behavioral source; the listed dependency trees explain referenced operators and algorithms only.

Submodule completeness is separate from the top-level source pins. NumPy's seven declared submodules were fetched at their recorded Git links. PyTorch's top-level tree and all 65 recursively declared submodules now match their recorded Git links. The interrupted third_party/cutlass checkout was moved to the repository's ignored build/source-recovery/ area before that one submodule was recreated at PyTorch's pinned commit; no other source tree was overwritten. The study sources remain outside the product build graph.

The PCM boundary is two distinct contracts: Kokoro's decoder and CustomSTFT define the generated waveform, while Android AAudio accepts a stream of 24 kHz mono float samples. The repository has a physical AAudio smoke result, but that did not execute the complete stock model. aosp-frameworks-av is study source only; no AAudio source was copied into the APK. The previously pinned AAudio commit and header blob are available as Git objects in that mirror and should be read with git show without changing its checkout.