Reply to thread

Message: [QUOTE="Bombastus, post: 92273, member: 2178"] Speech-separation systems can produce perceptually unnatural artifacts even when objective scores improve, so a clean-looking stem still needs a proper listening pass. Leakage is only one failure. Missing consonants, unstable tone, smeared ambience, and fragments assigned to the wrong speaker can be harder to notice. A separated voice may sound impressive in solo until you place it beside the original scene. Production dialogue carries room reflections, breaths, movement, microphone color, and tiny interactions between actors, so removing the competing speaker can disturb more than their words. Technical separation and usable dialogue are not automatically the same thing. Keep the untouched recording available throughout QC. Comparing processed audio only against itself makes gradual damage surprisingly easy to accept, especially after repeated listening. The original gives your ears a fixed reference for intelligibility, timing, tone, and room character. [HEADING=2]Separation errors hide in normal playback[/HEADING] Start at normal monitoring level with the separated stem against picture or the surrounding edit. Do not immediately hunt artifacts at exaggerated volume. The first pass should answer whether the actor still sounds like the actor and whether every intended word survives naturally inside the scene. Processing artifacts deserve attention even when the speaker has become easier to understand. [B][URL='https://doi.org/10.1109/TASLPRO.2026.3663910']Perceptual speech-separation quality[/URL][/B] can diverge from conventional objective measurements because processing may leave unnatural harmonic or speech structure behind. Your ears therefore remain part of the acceptance test, particularly for material headed into a finished mix. Listen around consonants first. Soft fricatives, plosives, word endings, and breaths carry relatively little energy and can be damaged before the center of a vowel sounds obviously wrong. A stem that retains the sentence but shaves detail from every quiet consonant can feel strangely blurred once music and effects arrive. Timbre can also wander inside a single phrase. One syllable may sound normal while the next becomes hollow, metallic, or oddly narrow because the separator had less certainty during the overlap. Short defects are easy to miss when your attention stays fixed on whether the unwanted speaker disappeared. [HEADING=2]Solo passes expose leakage and missing speech[/HEADING] Solo each separated output and follow the original dialogue line by line. Leakage often appears as recognizable syllables from the other speaker, but weaker remnants can resemble whispering, modulation, or room coloration. Headphones help here because low-level fragments can hide during ordinary speaker playback. Next, reverse the test. Listen specifically for pieces of the intended speaker that vanished from their own stem or appeared in the other output. Long-form speech-separation systems can suffer speaker-assignment problems, so a clean stem at the beginning does not guarantee the same voice stays correctly routed through every later overlap. Use [B][URL='https://goldmidi.com/community/threads/limited-edition-audio-has-introduced-voxsplit-pro-for-macos.77645/']speaker-separated dialogue stems[/URL][/B] as editable material rather than automatically treating either output as a finished replacement. A tiny leak under another actor's line may be harmless, while one missing consonant in an exposed close-up can make the entire render unusable. Context decides severity. Reverberant speech needs another pass. Room reflections continue after the direct voice, which means a separator has to deal with speech energy that extends beyond the obvious syllable. Listen to phrase endings and pauses for ambience that pumps, collapses, changes color, or seems to follow the wrong speaker. Do not ignore single-speaker sections surrounding the overlap. Some separation methods behave differently when only one person is active, and speech can be duplicated, thinned, or needlessly altered even though no separation was required there. Processing a shorter region can sometimes preserve more natural production sound than running an entire scene through the same treatment. [HEADING=2]The original mix decides what survives[/HEADING] Level-match the processed stem closely enough that louder does not automatically sound clearer. Then alternate between the original and separated version at the same scene position. Listen for meaning first, voice identity second, and production texture third, because a technically cleaner stem is useless if it changes a word or makes the actor sound detached from the location. Avoid judging quality by a null test alone. Modern separation can reconstruct or modify waveform details rather than simply subtracting one source from another, so the stems are not required to behave like traditional multitrack recordings that perfectly rebuild the mixture. Audible usefulness matters more than mathematical neatness. Use the original wherever separation causes more damage than the overlap itself. A clean production fragment can bridge into a processed section with short fades, allowing the separator to handle only the collision it actually solves. Longer processing regions create more opportunities for tone and ambience to drift. Check the repaired section once more with music, effects, and neighboring dialogue active. Some artifacts disappear harmlessly inside the mix, while missing speech detail becomes worse when other sounds mask it. The acceptable version is the one that preserves the performance with the least audible intervention, not necessarily the stem with the strongest isolation. [/QUOTE]

Name