wealy-asr-sbert-transformer-lyc

WEALY - wealy-asr-sbert-transformer-lyc

SBERT + Transformer model for version identification on LyricCovers dataset

This is a WEALY (WEakly-supervised Audio-LYrics) model for music version identification.

Usage

# Download and run inference
python scripts/inference.py \
    model_name=audio-based-lyrics-matching/wealy-asr-sbert-transformer-lyc \
    hidden_states=/path/to/your/hidden-states \
    partition=test \
    use_overlapping_chunks=true \
    ngpus=1

Required Path Arguments

When running inference, you must provide:

• hidden_states: Path to pre-extracted hidden states directory

Optional Arguments

• partition: Dataset partition to evaluate (default: "test")
• use_overlapping_chunks: Enable overlapping chunk evaluation (default: false)
• chunk_size: Size of overlapping chunks (default: 1500)
• overlap_percentage: Overlap between chunks (default: 0.9)
• ngpus: Number of GPUs to use (default: 1)

Model Details

This model was trained for version identification using the WEALY architecture.

Training Configuration

• Dataset: lyric-covers
• Embedding type: sbert
• Embedding dimension: 512

Citation

If you use this model, please cite:

@article{mancini2025wealy,
    title={Leveraging Whisper Embeddings for Audio-based Lyrics Matching},
    author={Mancini, Eleonora and Serrà, Joan and Torroni, Paolo and Mitsufuji, Yuki},
    journal={arXiv preprint arXiv:2510.08176},
    year={2025},
    url={https://github.com/helemanc/audio-based-lyrics-matching}
}