Inflect-Atom-v3 / THIRD_PARTY_NOTICES.md
owensong's picture
Stage verified private early-access package
9c1e420 verified
|
Raw History Blame Contribute Delete
1.26 kB

Provenance and Third-Party Notices

Inflect Atom is trained by distillation from synthetic speech/features produced by a voice-adapted Kokoro teacher. No Kokoro teacher checkpoint or learned neural vocoder is included or required at inference. This does not establish clearance of all model-weight, voice or training-data distribution rights; review remains pending. See BETA_ACCESS.md. No commercial-use permission is implied.

The small acoustic model and harmonic/noise DSP implementation are Inflect project code. The architectural design was informed by AuroraTTS-100K (AuroraAI-Research); no Aurora weights or runtime code are bundled.

Kokoro: https://github.com/hexgrad/kokoro (Apache-2.0). Misaki: https://github.com/hexgrad/misaki (separately installed). spaCy, PyTorch, NumPy, Numba, llvmlite and SoundFile are separately installed dependencies with their own licenses. The spaCy English model wheel is bundled in vendor/ and contains its own license files. The pronunciation fallback requires separately installed phonemizer-fork, espeakng-loader and eSpeak-ng; upstream GPL obligations apply independently. No eSpeak binaries or eSpeak dictionary data are bundled here. The supplied software notices do not grant a final license for Atom weights.