01 / AI · Analysis
FoleyForge proposes sound editing that preserves identity
The design combines a training pipeline that creates editing examples with controls for loudness and edit strength.

The FoleyForge paper proposed a single model that edits sound effects from an input clip and a written instruction while preserving the original sound's identity.
The paper identified a shortage of publicly available audio editing data as a development challenge. Its proposed approach connects the creation of training examples with controls over how the model applies an edit.
The editing scope covers changes to meaning and style, alongside modifications resembling digital signal processing—the numerical manipulation of audio. The proposed design assigns these different types of edits to the same model.
Creating editing examples for training
The paper proposed guiding an audio language model to produce editing examples. The pipeline generates both acoustic edits based on digital signal processing and changes to meaning that retain similarities to the original sound.
The next stage uses those generated examples to further train an audio generation model. This makes examples of sound-effect editing part of the model's training process.
Separate controls for edit strength
The paper describes controls for loudness change and edit strength. It also includes a control based on contour information extracted from the input sound.
The paper reported instruction-following performance, audio quality and faithfulness to the source as outcomes of the approach. These are author-reported results; the abstract does not disclose performance figures for individual comparison models or the evaluation conditions.