Readme
Model Details
MAEST is a family of Transformer models based on PASST and focused on music analysis applications. The MAEST models are also available for inference in the Essentia library and for inference and training in the official repository .
Model Description
- Developed by: Pablo Alonso
- Shared by: Pablo Alonso
- Model type: Transformer
- License: cc-by-nc-sa-4.0
- Finetuned from model: PaSST
Model Sources
- Repository: MAEST
- Paper: Efficient Supervised Training of Audio Transformers for Music Representation Learning
Uses
MAEST is a music audio representation model pre-trained on the task of music style classification. According to the evaluation reported in the original paper, it reports good performance in several downstream music analysis tasks.
Direct Use
The MAEST models can make predictions for a taxonomy of 400 music styles derived from the public metadata of Discogs .
Downstream Use
The MAEST models have reported good performance in downstream applications related to music genre recognition, music emotion recognition, and instrument detection. Specifically, the original paper reports that the best performance is obtained from representations extracted from intermediate layers of the model.
Out-of-Scope Use
The model has not been evaluated outside the context of music understanding applications, so we are unaware of its performance outside its intended domain.
Since the model is intended to be used within the
audio-classification
pipeline, it is important to mention that MAEST is
NOT
a general-purpose audio classification model (such as
AST
), so it shuold not be expected to perform well in tasks such as
AudioSet
.
Bias, Risks, and Limitations
The MAEST models were trained using Discogs20, an in-house MTG dataset derived from the public Discogs metadata. While we tried to maximize the diversity with respect to the 400 music styles covered in the dataset, we noted an overrepresentation of Western (particularly electronic) music.
Training Details
Training Data
Our models were trained using Discogs20, MTG in-house dataset featuring 3.3M music tracks matched to Discogs ’ metadata.
Training Procedure
Most training details are detailed in the paper and official implementation of the model.
Preprocessing
MAEST models rely on mel-spectrograms originally extracted with the Essentia library, and used in several previous publications.
In Transformers, this mel-spectrogram signature is replicated to a certain extent using
audio_utils
, which have a very small (but not neglectable) impact on the predictions.
Evaluation, Metrics, and results
The MAEST models were pre-trained in the task of music style classification, and their internal representations were evaluated via downstream MLP probes in several benchmark music understanding tasks. Check the original paper for details.
Environmental Impact
- Hardware Type: 4 x Nvidia RTX 2080 Ti
- Hours used: apprx. 32
- Carbon Emitted: apprx. 3.46 kg CO2 eq.
Carbon emissions estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019) .
Technical Specifications
Model Architecture and Objective
Audio Spectrogram Transformer (AST)
Compute Infrastructure
Local infrastructure
Hardware
4 x Nvidia RTX 2080 Ti
Software
Pytorch
Citation
BibTeX:
@inproceedings{alonso2023music,
title={Efficient supervised training of audio transformers for music representation learning},
author={Alonso-Jim{\'e}nez, Pablo and Serra, Xavier and Bogdanov, Dmitry},
booktitle={Proceedings of the 24th International Society for Music Information Retrieval Conference (ISMIR 2023)},
year={2022},
organization={International Society for Music Information Retrieval (ISMIR)}
}
APA:
Alonso-Jiménez, P., Serra, X., & Bogdanov, D. (2023). Efficient Supervised Training of Audio Transformers for Music Representation Learning. In Proceedings of the 24th International Society for Music Information Retrieval Conference (ISMIR 2023)
Model Card Authors
Pablo Alonso
Model Card Contact
-
Twitter: @pablo__alonso
-
Github: @palonso
-
mail: pablo
dotalonsoatupfdotedu