Abstract

This paper discloses attribution rules, provenance-marking requirements, a training-objective constraint and a model-release restriction for generative voice and style models trained on the recordings of deceased artists whose estates, rights-holders and records of intent do not survive. The worked case is the Cambodian musicians and filmmakers killed under the Khmer Rouge (1975–1979), where consent cannot be obtained, no rights-holding institution survives, and the destruction was aimed at the culture, so that an indistinguishable reconstruction completes the erasure. Generated work is credited to a named living human creator and published for, never by, the deceased artist: a dedicated work, not an authorship claim. Each medium takes its own operation: preservation for film and plays, offering for music, new authorship for story, invention for scripture. Audio restoration and source separation are the data-preparation stage for voice-model training, so generation must stay downstream of restoration and be governed separately. Beyond fitting the model to the source recordings, the training and selection objective admits only a refusal signal and excludes revenue, engagement and all other audience-response metrics. Provenance marking is permanent and its obligation rises with output realism; lost or unfinished works are never synthesised; revenue routes to heirs, living practitioners and restoration rather than to a manufactured rights-holder; heir consent is sought before any generation wherever an heir can be found, and an heir's refusal, even a later one, stops the work; and model weights are never released. Stated limits: nothing is built, heirs are largely unidentifiable, and model leakage is unsolved.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS