docs01
Audio processing chain
The MossFormer2 neural engine runs locally at 48 kHz with Apple MLX hardware acceleration. Longer files are chunked into 60-second segments with overlap context to avoid seam artifacts.
01Speech Isolation
02Original ambience mix
03Vocal EQ & De-esser
04−14 LUFS Mastering
02
Balanced natural ambience
Transcribed speech regions preserve a fraction of the natural stereo room tone so vocals never sound artificial or robotic. In pauses, background noise is suppressed cleanly.
03
Vocal profiles & A/B audition
Use the instant A/B audition tool to listen to the exact same sentence before and after Pure Voice processing.
ProfileCharacterRecommended for
NaturalSubtle clean-upQuiet studios, natural conversations
CleanVersatile & balancedStandard video voiceovers and desk mics
StudioMaximum isolationEchoey rooms, HVAC rumble, and noisy backgrounds
04
Recalculation and caching
Neural inference results are cached independently from EQ and mastering settings. Toggling Pure Voice on or off never forces a redundant model re-run.