The Local Vocal Remover is an offline audio processing tool that uses an on-device machine learning model to remove the vocal track from music recordings. Unlike the Vocal and Music Separator which produces both vocal and instrumental outputs, the Local Vocal Remover is optimised specifically for producing a clean instrumental with minimal vocal residue.
The tool is designed for users who need karaoke-style instrumentals, background music for video projects, or instrumental versions of songs for practice or arrangement study. Because it runs entirely offline after the model weights are downloaded on the first use, it is suitable for use without an internet connection and does not transmit audio data anywhere.
The separation technique uses a neural network trained on a large dataset of music recordings. The model has learned the acoustic signature of human voices — the frequency patterns, the temporal envelope, the formant structure — and is able to estimate a mask that suppresses those components in the output. The underlying approach is similar to the short-time Fourier transform masking technique, enhanced with learned parameters.
The limitation of any vocal removal tool is that instruments and vocals often share frequency ranges. A piano playing a note at the same pitch as a sung note will produce energy at the same frequency. The model must decide which energy belongs to the vocal and which to the instrument, and it cannot always get this right. The result is an instrumental that may have some vocal ghosting, particularly during melodic passages where pitch alignment between voice and instrument is common.
Despite these limitations, the output is useful for a wide range of practical purposes. For karaoke use, a slight vocal ghost is acceptable because the performer's own voice dominates the listening mix. For background music in video, the vocal reduction is typically sufficient. For detailed production work, the output can be taken into an audio editor for further processing.
Local Vocal Remover is part of the Envizion Tools2 collection at envizion.work.