Google Inc. a database of reference mixed sounds, equipped with annotations, that can be used in machine learning systems for separating arbitrary mixed sounds into distinct components. A universal deep learning model (TDCN++) has also been published, which can be used in Tensorflow for sound separation. The data has been prepared based on a collection and under the CC BY 4.0 license.
The FUSS (Free Universal Sound Separation) project aims to solve the problem of separating any number of arbitrary sounds, the nature of which is not known in advance. Other similar systems are usually limited to the task of separating specific sounds, such as voice and non-voice or different speakers.
The database contains approximately 20,000 mixtures. The set also includes pre-calculated impulse responses of the room, prepared using a specially created room simulator, taking into account reflections from walls, the position of the sound source, and the position of the microphone.
Source: opennet.ru
