Reply to thread

Message: [QUOTE="Bombastus, post: 91175, member: 2178"] Article 50(2) exempts systems used for standard editing and systems that do not substantially alter supplied audio or its meaning. That exception matters to producers using noise removal, stem separation, pitch correction, or automated mastering, because not every AI-assisted file needs a machine-readable mark. The duty in this part of the law belongs to the provider of the system that generates or manipulates synthetic content. It is not a blanket order for every musician or engineer to label every session touched by an algorithm. The user’s choices still matter because the provider must know what its system actually produces. The practical dividing line is creation, not the marketing label on the plug-in. A tool that clarifies a performance you recorded sits differently from one that invents a singer, replaces words, or supplies notes nobody played. [HEADING=2]Routine cleanup remains ordinary editing[/HEADING] Noise reduction is the cleanest example. Removing steady hiss, clicks, hum, or room noise normally improves access to the captured performance without changing who performed, what was said, or which musical event occurred. That looks like [B]routine production cleanup[/B], even when the software learned the repair from data. The law gives providers two routes to the exception. An operation may be standard editing, or it may simply leave the input and its semantics substantially unchanged, which matters when an unfamiliar repair tool does not fit a conventional studio category. For audio, meaning reaches beyond the dictionary sense of speech. Changing the speaker’s identity, the words delivered, or the apparent existence of a performance can alter what the recording asserts even when its loudness and duration barely move. Level matching, de-essing, modest EQ, and restrained pitch correction can fall on the same side when they polish existing material instead of creating new content. Automated mastering is not automatically exempt, though. A process marketed as mastering may still synthesize drums, extend a passage, or rewrite a vocal. Stem separation needs the same functional test. Extracting a vocal already present in the mix is closer to processing existing input, while regenerating missing frequencies or fabricating an isolated part can produce materially new audio. The feature name proves nothing. Automatic transcription, subtitles, and text-to-speech used for accessibility generally count as standard processing in ordinary cases. That does not turn a commercial synthetic narrator into an accessibility tool merely because listeners can hear it more clearly. [HEADING=2]New performances cross the line quickly[/HEADING] Voice replacement is the obvious break. If a system substitutes a different speaker, generates a sentence the person never recorded, or creates a sung phrase from text, it has moved beyond repair into synthetic performance. The provider-side marking obligation can attach even when nobody is impersonated. That last distinction is often missed. Machine-readable marking under Article 50(2) covers synthetic audio broadly, while a listener-facing disclosure under Article 50(4) is tied to content that meets the deepfake test. One technical duty does not expand the other. Your release therefore needs two separate decisions. First decide whether the system provider had to mark the output, then apply [B][URL='https://goldmidi.com/community/threads/the-eu-labeling-rules-that-now-apply-to-ai-audio.76573/']the EU rules for synthetic audio disclosure[/URL][/B] to the way the finished recording represents people, objects, places, or events. A watermark cannot answer the second question for the listener. Responsibility stays at the AI-system layer even when a plug-in developer relies on an upstream model or a third-party marking service. The developer may reuse that technology, but it must still be able to show that detection is effective, interoperable, robust, and reliable within current technical limits. [HEADING=2]Finished audio still needs a second legal test[/HEADING] An editing exemption only answers whether Article 50(2) requires provider-side marking. It does not decide whether a professional deployer must disclose a realistic voice clone at first exposure, and it does not erase accessibility requirements for that disclosure. Different duties, different actors. Keep the classification tied to the actual operation. Record the original input, the tool and version, what the system changed, whether it generated new speech or music, and why the change preserved or altered the source’s substance. A vague project note saying “AI assisted” cannot show where the boundary sat. This record becomes particularly useful when an apparently modest edit moves between tools. Noise cleanup may remain ordinary, then a later voice-conversion stage may create a new performance that needs marking. Treat the chain as separate operations rather than letting the harmless first step hide the consequential second one. The marking analysis also leaves consent, copyright, performer rights, and contract terms untouched. An unmarked edit can still be unlawful for another reason, while a properly marked synthetic vocal can still breach a voice license. The file’s transparency status never proves permission. [/QUOTE]

Name