laion/NotebookLM-Podcasts-Humaneness-Prepared

Five-language annotated NotebookLM podcasts — upload in progress

This is the selected English, German, French, Spanish and Dutch derivative of TTS-AGI/NotebookLM-Podcasts. The selected source inventory has 7,158 episodes, each passing both beginning and middle DNSMOS background-quality >=4.0. Source language labels are publisher metadata, not independently verified language. Other source languages are not part of this release and are not claimed complete here.

The packed TAR upload is NOT complete yet. A release is complete only after PUBLICATION_COMPLETE.json exists and all remote TAR checksums have been verified. The final package preserves original audio, episode-global word timestamps, speaker diarization/activity, all 192 Humaneness Ears medium scores, sentence captions, burst events and continuous MOSS-v2 timelines with all 32 codebooks. The current model uses the first 12 codebooks at 12.5 frames/s.

The source publisher labels recordings Public Domain; preserve that attribution without treating it as an independent rights audit. LAION-authored annotations and documentation are CC BY4.0; upstream dependencies retain their own terms.

The separate three-epoch full fine-tune starts from final Humaneness Voice Large 1.3 Alpha, not original SFT3. It learns speaker-labeled identity references and past conversation context. The public listening page is initially marked in progress and will receive 50 held-out predictions after the final training checkpoint passes completion gates.