Streaming provider Deezer the source code of the experimental project Spleeter, which develops a machine learning system for separating audio sources from complex sound compositions. The program allows for the removal of vocals and retains only the musical accompaniment, manipulates the sound of individual instruments, or discards the music to leave the voice for overlaying onto another audio track, creating mixes, karaoke, or transcription. The project code is written in Python using the Tensorflow engine and pod MIT licencom.
Za preuzimanje already trained models for separating vocals (a single voice) from the accompaniment, as well as for separating into 4 and 5 streams, including vocals, drums, bass, piano, and other sounds. Spleeter can be used both as a Python library and as a standalone command-line utility. In its simplest case, based on the source file two, four, or five files with vocals and parts from the accompaniment (vocals.wav, drums.wav, bass.wav, piano.wav, other.wav).
When separating into 2 and 4 streams, Spleeter provides very high performance; for example, when using a GPU, separating an audio file into 4 streams takes 100 times less time than the duration of the original composition. On a system with an NVIDIA GeForce GTX 1080 GPU and a 32-core Intel Xeon Gold 6134 CPU, processing a test collection from musDB, lasting three hours and 27 minutes, was completed in 90 seconds.
Among the advantages of Spleeter, compared to other developments in the field of sound separation, such as the open project , is mentioned the application of higher quality models built on the basis of an extensive collection of audio files. Due to copyright restrictions, researchers in machine learning are limited to access to rather scant publicly available collections of musical files, while for Spleeter, models were built using data from Deezer's extensive music catalog.
Po s Open-Unmix alat Spleeter vrši razdvajanje otprilike 35% brže kada se testira na CPU, podržava MP3 datoteke i generira znatno kvalitetniji rezultat (pri izdvajanju vokala u Open-Unmix ostaju tragovi nekih instrumenata, što vjerojatno objašnjava činjenica da su modeli Open-Unmix trenirani na kolekciji od ukupno 150 kompozicija).
Izvor: opennet.ru
