Fixing Artefacts
Tell a generation artefact from a performance choice, then decide whether to repair or regenerate.
Related
The problem
A generation returns clicks, blurred tails, drifting pitch and parts that sit behind a sheet of water. None of this proves the model failed. The prompt, the length, the density or the reference may simply have pushed it to an edge. Identify what the fault is before deciding whether it deserves repair.
How it works
Artefacts have recognisable sources. A click at an attack usually comes from a splice, a short envelope or a cut sample. A blurred sustain comes from reverb tails, release handling or too much overlap between notes. Unstable pitch appears in voice and sustained strings, where the model has to invent vibrato and portamento. Watery separation comes from a separation model leaving bleed between parts. A long generation also risks seams where the model splices at a breath, and those seams show up as jumps in level. Performance produces similar effects, since vibrato, portamento, pedal and room sound all reduce cleanliness; that overlap is what makes the judgement hard.
Practice
Mute reverb and compression and listen dry, then locate the fault in a single layer. Small faults yield to trimming a few milliseconds before an attack, drawing a short fade, or correcting one note. When a whole passage is blurred, parts bleed into each other, or the timing has slipped, regeneration is cheaper than repair. Change one condition per regeneration, such as a shorter clip, lower density or a simpler texture, never five at once. Keep the previous version until the new one has been compared against it, and put the fix back into the whole to check that the phrase still breathes.
What to listen for
On the funk kit, whether attacks stay crisp; on the organ hymn, whether long tails blur; on Le gibet, whether the ostinato and the chords remain separate; on the Mahler song, whether the voice holds pitch without unintended wobble. Comparing the four slowly draws the line between artefact and performance.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Sources
- 生成式音频常见伪影的类别与成因,依公开技术资料归纳,未抄原文
- 修复与重生成的取舍,依后期制作通识核对,未抄文本
Used for fact-checking only; all prose is original.