yuzu/spleeter
1
0
Fork 0
mirror of https://github.com/YuzuZensai/spleeter.git synced 2026-01-06 04:32:43 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #384 from faroit/joss_proofing

Browse Source 
[JOSS] paper proofing
This commit is contained in:
Romain Hennequin
2020-05-25 16:56:22 +02:00
committed by GitHub
parent abda77661b b1cff3ccb9
commit b348f18b65
1 changed files with 14 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
paper.md
View File

@@ -1,5 +1,5 @@
---
title: 'Spleeter: a fast and state-of-the art music source separation tool with pre-trained models'
title: 'Spleeter: a fast and state-of-the-art music source separation tool with pre-trained models'
tags:
- Python
- musical signal processing
@@ -26,19 +26,19 @@ bibliography: paper.bib
## Summary
We present and release a new tool for music source separation with pre-trained models called Spleeter. Spleeter was designed with ease of use, separation performance and speed in mind. Spleeter is based on Tensorflow [@tensorflow2015-whitepaper] and makes it possible to:
We present and release a new tool for music source separation with pre-trained models called Spleeter. Spleeter was designed with ease of use, separation performance, and speed in mind. Spleeter is based on Tensorflow [@tensorflow2015-whitepaper] and makes it possible to:
- split music audio files into several stems with a single command line using pre-trained models. A music audio file can be separated into $2$ stems (vocals and accompaniments), $4$ stems (vocals, drums, bass and other) or $5$ stems (vocals, drums, bass, piano and other).
- split music audio files into several stems with a single command line using pre-trained models. A music audio file can be separated into $2$ stems (vocals and accompaniments), $4$ stems (vocals, drums, bass, and other) or $5$ stems (vocals, drums, bass, piano and other).
- train source separation models or fine-tune pre-trained ones with Tensorflow (provided you have a dataset of isolated sources).
The performance of the pre-trained models are very close to the published state of the art and is one of the best performing $4$ stems separation model on the common musdb18 benchmark [@musdb18] to be publicly released. Spleeter is also very fast as it can separate a mix audio file into $4$ stems $100$ times faster than real-time (we note, though, that the model cannot be applied in real-time as it needs buffering) on a single Graphics Processing Unit (GPU) using the pre-trained $4$-stems model.
The performance of the pre-trained models are very close to the published state-of-the-art and is one of the best performing $4$ stems separation model on the common musdb18 benchmark [@musdb18] to be publicly released. Spleeter is also very fast as it can separate a mix audio file into $4$ stems $100$ times faster than real-time (we note, though, that the model cannot be applied in real-time as it needs buffering) on a single Graphics Processing Unit (GPU) using the pre-trained $4$-stems model.
## Purpose
We release Spleeter with pre-trained state-of-the-art models in order to help the Music Information Retrieval (MIR) research community leverage the power of source separation in various MIR tasks, such as vocal lyrics analysis from audio (audio/lyrics alignement, lyrics transcription...), music transcription (chord transcription, drums transcription, bass transcription, chord estimation, beat tracking), singer identification, any type of multilabel classification (mood/genre...), vocal melody extraction or cover detection.
We believe that source separation has reached a level of maturity that makes it worth consideration for these tasks and that specific features computed from isolated vocals, drums or bass may help increase performances, especially in low data availability scenarios (small datasets, limited annotation availability) for which supervised learning might be difficult.
Spleeter also makes it possible to fine tune the provided state-of-the-art models in order to adapt the system to a specific use-case.
Finally, having an available source separation tool such as Spleeter will allow researchers to compare performances of their new models to a state-of-the-art one on their own private datasets instead of musdb18, which is usually the only used dataset for reporting separation performances for unreleased models.
We release Spleeter with pre-trained state-of-the-art models in order to help the Music Information Retrieval (MIR) research community leverage the power of source separation in various MIR tasks, such as vocal lyrics analysis from audio (audio/lyrics alignment, lyrics transcription...), music transcription (chord transcription, drums transcription, bass transcription, chord estimation, beat tracking), singer identification, any type of multilabel classification (mood/genre...), vocal melody extraction or cover detection.
We believe that source separation has reached a level of maturity that makes it worth considering for these tasks and that specific features computed from isolated vocals, drums or bass may help increase performances, especially in low data availability scenarios (small datasets, limited annotation availability) for which supervised learning might be difficult.
Spleeter also makes it possible to fine-tune the provided state-of-the-art models in order to adapt the system to a specific use-case.
Finally, having an available source separation tool such as Spleeter will allow researchers to compare performances of their new models to a state-of-the-art one on their private datasets instead of musdb18, which is usually the only used dataset for reporting separation performances for unreleased models.
Note that we cannot release the training data for copyright reasons, and thus, sharing pre-trained models were the only way to make these results available to the community.
## Implementation details
@@ -47,11 +47,11 @@ Spleeter contains pre-trained models for:
- vocals/accompaniment separation.
- $4$ stems separation as in SiSec [@SISEC18] (vocals, bass, drums and other).
- $5$ stems separation with an extra piano stem (vocals, bass, drums, piano and other). It is, to the authors knowledge, the first released model to perform such a separation.
- $5$ stems separation with an extra piano stem (vocals, bass, drums, piano, and other). It is, to the authors' knowledge, the first released model to perform such a separation.
The pre-trained models are U-nets [@unet2017] and follows similar specifications as in [@deezerICASSP2019]. The U-net is a encoder/decoder Convolutional Neural Network (CNN) architecture with skip connections. We used $12$-layer U-nets ($6$ layers for the encoder and $6$ for the decoder). A U-net is used for estimating a soft mask for each source (stem). Training loss is a $L_1$-norm between masked input mix spectrograms and source target spectrograms. The models were trained on Deezer internal datasets (noteworthily the Bean dataset that was used in [@deezerICASSP2019]) using Adam [@Adam]. Training time took approximately a full week on a single GPU. Separation is then done from estimated source spectrograms using soft masking or multi-channel Wiener filtering.
The pre-trained models are U-nets [@unet2017] and follow similar specifications as in [@deezerICASSP2019]. The U-net is an encoder/decoder Convolutional Neural Network (CNN) architecture with skip connections. We used $12$-layer U-nets ($6$ layers for the encoder and $6$ for the decoder). A U-net is used for estimating a soft mask for each source (stem). Training loss is a $L_1$-norm between masked input mix spectrograms and source-target spectrograms. The models were trained on Deezer's internal datasets (noteworthily the Bean dataset that was used in [@deezerICASSP2019]) using Adam [@Adam]. Training time took approximately a full week on a single GPU. Separation is then done from estimated source spectrograms using soft masking or multi-channel Wiener filtering.
Training and inference is implemented in Tensorflow which makes it possible to run the code on Central Processing Unit (CPU) or GPU.
Training and inference are implemented in Tensorflow which makes it possible to run the code on Central Processing Unit (CPU) or GPU.
## Speed
@@ -59,7 +59,7 @@ As the whole separation pipeline can be run on a GPU and the model is based on a
## Separation performances
The models compete with the state of the art on the standard musdb18 dataset [@musdb18] while it was not trained, validated or optimized in any way with musdb18 data. We report results in terms of standard source separation metrics [@separation_metrics], namely Signal to Distortion Ratio (SDR), Signal to Artifacts Ratio (SAR), Signal to Interference Ratio (SIR) and source Image to Spatial distortion Ratio (ISR), are presented in the following table compared to Open-Unmix [@Open-Unmix] and Demucs [@demucs] (only SDR are reported for Demucs since other metrics are not available in the paper) which are, to the authors knowledge, the only released system that perform near state-of-the-art performances.
The models compete with the state-of-the-art on the standard musdb18 dataset [@musdb18] while it was not trained, validated or optimized in any way with musdb18 data. We report results in terms of standard source separation metrics [@separation_metrics], namely Signal to Distortion Ratio (SDR), Signal to Artifacts Ratio (SAR), Signal to Interference Ratio (SIR) and source Image to Spatial distortion Ratio (ISR), are presented in the following table compared to Open-Unmix [@Open-Unmix] and Demucs [@demucs] (only SDR are reported for Demucs since other metrics are not available in the paper) which are, to the authors' knowledge, the only released system that performs near state-of-the-art performances.
We present results for soft masking and for multi-channel Wiener filtering (applied using Norbert [@Norbert]). As can be seen, for most metrics Spleeter is competitive with Open-Unmix and especially on SDR for all instruments, and is almost on par with Demucs.
@@ -82,12 +82,11 @@ We present results for soft masking and for multi-channel Wiener filtering (appl
| Other SAR |4.63 |4.88 |4.74 |5.26 |
| Other ISR |9.83 |9.87 |9.31 |10.86 |
Spleeter [@spleeter] source code and pre-trained models are available on [github](https://www.github.com/deezer/spleeter) and distributed under a MIT license. This repository will eventually be used for releasing other models with improved performances or models separating into more than $5$ stems in the future.
Spleeter [@spleeter] source code and pre-trained models are available on [github](https://www.github.com/deezer/spleeter) and distributed under a MIT license. This repository will eventually be used for releasing other models with improved performances or models separating into more than $5$ stems in the future.
## Distribution
Spleeter is available as a standalone Python package, and also provided as a [conda](https://github.com/conda-forge/spleeter-feedstock) recipe and self-contained [Dockers](https://hub.docker.com/r/researchdeezer/spleeter) which makes it usable as is on various platforms.
Spleeter is available as a standalone Python package, and also provided as a [conda](https://github.com/conda-forge/spleeter-feedstock) recipe and self-contained [Dockers](https://hub.docker.com/r/researchdeezer/spleeter) which makes it usable as-is on various platforms.
## Acknowledgements