Audio Source Separation

Audio source separation estimates the individual sounds hidden in a recording, supporting clearer speech, editable music stems, and selective control of noisy soundscapes. Methods range from spatial filtering and nonnegative matrix factorization to neural mask estimators such as Demucs and Band-Split RNN, with newer Transformer, flow-matching, and diffusion models guided by text, video, speaker identity, or musical scores. Systems are tested on speech mixtures such as Libri2Mix, music benchmarks such as MUSDB18, and open-domain sound datasets, using signal metrics alongside listening tests and downstream tasks. A central challenge is balancing separation quality and perceptual naturalness with generalization, real-time efficiency, and reliable performance when separated audio is used for tasks such as speech recognition.