| Paradis-TTS — provenance notices | |
| paradis_tts/onnx_models.py adapts the single-sentence text forward and | |
| convolutional STFT from Paradee training/scripts/export_paradee.py, | |
| Sahil Mahendrakar, commit 9c8b4d7504cbee7e64de2d0341bb690f0b0ab708, | |
| Apache-2.0. Changes: French A_full_phase, three explicit deterministic ONNX | |
| stages, host excitation and host postprocessing, CPU parity checks and | |
| portable FP32/INT8-weight bundles. See LICENSE and docs/PHASE9_ONNX.md. | |
| The portable NumPy phase filter in paradis_tts/audio.py implements the same | |
| Paradee relative-phase correction described below; existing training sources | |
| and their licenses are unchanged. | |
| Paradis-TTS — phase 0 French feasibility audit | |
| The teacher extraction in scripts/test_teacher_fr.py is adapted from: | |
| Paradee, Sahil Mahendrakar | |
| https://github.com/sahilmahendrakar/paradee | |
| Commit: 9c8b4d7504cbee7e64de2d0341bb690f0b0ab708 | |
| File: training/scripts/gen_teacher.py | |
| License: Apache-2.0 (see LICENSE) | |
| Changes: French ff_siwis, pinned assets, CPU determinism, strict vocabulary | |
| validation, hooks for independent parity checks, logs and test artifacts. | |
| Kokoro-82M and Kokoro / Misaki software: hexgrad, Apache-2.0. | |
| https://huggingface.co/hexgrad/Kokoro-82M | |
| https://github.com/hexgrad/kokoro | |
| https://github.com/hexgrad/misaki | |
| French teacher voice: ff_siwis. Kokoro attributes its source data to: | |
| Yamagishi, Junichi; Honnet, Pierre-Edouard; Garner, Philip; Lazaridis, Alexandros. | |
| (2017). The SIWIS French Speech Synthesis Database, 2016 [dataset]. | |
| University of Edinburgh, School of Informatics, Centre for Speech Technology Research. | |
| https://doi.org/10.7488/ds/1705 | |
| Source-data license: https://creativecommons.org/licenses/by/4.0/ | |
| The included WAVs are synthetic Kokoro outputs from original test sentences; | |
| they are not recordings downloaded from SIWIS. | |
| No WikiText or Wikipedia text is redistributed in phase 0 evidence. | |
| Phase 1 includes French Wikipedia sentence excerpts in data/: | |
| Contributors to French Wikipedia; article and history URLs in metadata.jsonl. | |
| Source: wikimedia/wikipedia, 20231101.fr, | |
| revision b04c8d1ceb2f5cd4588862100d08de323dccfbaa. | |
| Original source-declared license: CC BY-SA 3.0. | |
| Adapted corpus / original coverage sentences: CC BY-SA 4.0. | |
| See data/README.md and docs/LICENSING.md for notices and changes. | |
| Phase 2 extraction adapts Paradee training/scripts/gen_teacher.py and | |
| training/scripts/regen_audio.py at the same audited Paradee commit. | |
| Changes include French corpus/voice, stable per-ID seeds, atomic tensor shards, | |
| checked resume, and a CUDA decoder with original CPU excitation and FFT math. | |
| Generated teacher audio is kept in tensor shards without additional WAV exports. | |
| Dependency licenses, including eSpeak NG's GPL, remain separate from this | |
| repository's source-code license; see docs/LICENSING.md. | |
| Phases 3/4 include unmodified Paradee training/scripts/student.py and | |
| student_decoder.py in training/vendor/, under Apache-2.0 at the audited commit. | |
| training/batches.py and losses.py adapt its baseline collators and loss formulas. | |
| Changes: French split-aware smoke subsets, deterministic CPU reflection padding, | |
| small batches/accumulation, portable RNG/optimizer/scaler checkpoints and Colab tools. | |
| No English Paradee model weights or English corpus are used for French training. | |
| Phase 6 includes unmodified Paradee training/scripts/discriminators.py at the | |
| same audited commit. training/decoder_training.py and decoder_gan.py adapt | |
| train_decoder.py: French split-aware data, spectral/GAN stages, CPU reflection | |
| padding, exact global STFT convergence across accumulated microbatches, full | |
| optimizer/discriminator/RNG resume, Drive backups and tensor/HTML listening. | |
| paradis_tts/phase_lock.py adapts Paradee training/scripts/phase_lock.py at the | |
| same audited Apache-2.0 commit. It is initially exposed in a CPU listening A/B | |
| test; the raw student inference and trained checkpoints remain the references. | |
| tests/fixtures/upstream_phase_lock.py contains the original smooth/lock | |
| function bodies for a standalone parity test, under the same Apache-2.0 license. | |
| The optional high-frequency loss in training/decoder_hf.py adapts the | |
| --hf-weight objective from Paradee training/scripts/train_decoder.py at | |
| commit 9c8b4d7504cbee7e64de2d0341bb690f0b0ab708 (Apache-2.0). | |
| Evaluation uses the non-personalized DNSMOS P.835 windowing/calibration method | |
| from Microsoft DNS-Challenge, revision 591184a9fcb2cbdec02520fed81a32bbbf9d73ff, | |
| DNSMOS/dnsmos_local.py. Source repository license: CC BY 4.0. | |
| https://github.com/microsoft/DNS-Challenge | |
| https://creativecommons.org/licenses/by/4.0/ | |
| Adaptation: P.835 only (no P.808), SciPy polyphase resampling, explicit CPU thread | |
| limits, resumable per-sentence logs and integration with held-out TTS evaluation. | |
| DNSMOS model files are downloaded for evaluation and are not redistributed in | |
| the Python package or inference archives. This attribution does not imply | |
| Microsoft endorses Paradis-TTS. | |