AI Music Glossary

A class of machine learning models that learn the statistical patterns of a training dataset and use those patterns to produce new content — audio, images, text, video — that resembles but is not identical to the training data. In music, generative AI models can produce melodies, harmonies, rhythms, vocal performances, and complete arrangements from text prompts or musical inputs.

A generative AI architecture that learns to create content by first encoding data (audio, images) into a compressed 'latent space', then training a diffusion model to generate realistic outputs by gradually denoising random noise in that space. For audio, the model encodes waveforms into spectrograms or audio tokens, learns the patterns of music in that compressed representation, and generates new audio by starting from noise and iteratively refining it toward a coherent output. Used in tools including Stable Audio (Stability AI) and several research systems.

An AI audio processing technique that isolates individual instrument or vocal tracks from a finished stereo mix — without access to the original multitrack session. Models are trained on large collections of paired mixed audio and individual stems, learning to predict where each source's audio energy sits within the combined signal. Output quality has improved dramatically since 2019; modern tools like Meta's Demucs can separate 6 stems with acceptable bleed in most genres.

The use of machine learning algorithms to apply mastering processing — EQ, compression, limiting, stereo widening, loudness normalisation — to an unmastered audio mix, without human engineer intervention. AI mastering services analyse the spectral and dynamic characteristics of the submitted mix and compare them against a learned model of commercially successful masters, then apply corrective processing to match. Services like LANDR, eMastered, and Matchering use different approaches: LANDR uses its own trained model; Matchering explicitly matches the profile of a reference track you specify.

The dataset of existing content — recordings, scores, MIDI files, audio samples — that an AI model learns from before it can generate new music. The quality, diversity, and licensing status of training data profoundly shapes what a model can produce and what legal liabilities attach to using it. A model trained primarily on Western pop music will reflect those aesthetic norms; one trained on a diverse global corpus will be more stylistically flexible. The question of whether training on copyrighted recordings without a license constitutes infringement is the central legal dispute in AI music as of 2025.