Jam-ALT: A Readability-Aware Lyrics Transcription Benchmark

Jam-ALT is a revision of the JamendoLyrics dataset (79 songs in 4 languages), intended for use as an automatic lyrics transcription (ALT) benchmark. It has been published in the ISMIR 2024 paper (full citation below):
📄 Lyrics Transcription for Humans: A Readability-Aware Benchmark
👥 O. Cífka, H. Schreiber, L. Miner, F.-R. Stöter
🏢 AudioShake

The lyrics have been revised according to the newly compiled annotation guidelines, which include rules about spelling and formatting, as well as punctuation and capitalization (PnC). The audio is identical to the JamendoLyrics dataset.

Note: The dataset is not time-aligned as it does not easily map to the timestamps from JamendoLyrics. To evaluate automatic lyrics alignment (ALA), please use JamendoLyrics directly.

See the project website for details and the JamendoLyrics community for related dataset.

Loading the data

from datasets import load_dataset
dataset = load_dataset("audioshake/jam-alt", split="test")

A subset is defined for each language (en, fr, de, es); for example, use load_dataset("audioshake/jam-alt", "es") to load only the Spanish songs.

To control how the audio is decoded, cast the audio column using dataset.cast_column("audio", datasets.Audio(...)). Useful arguments to datasets.Audio() are:

• sampling_rate and mono=True to control the sampling rate and number of channels.
• decode=False to skip decoding the audio and just get the MP3 file paths and contents.

The load_dataset function also accepts a columns parameter, which can be useful for example if you want to skip downloading the audio (see the example below).

Running the benchmark

The evaluation is implemented in our alt-eval package:

from datasets import load_dataset
from alt_eval import compute_metrics

dataset = load_dataset("audioshake/jam-alt", revision="v1.2.0", split="test")
# transcriptions: list[str]
compute_metrics(dataset["text"], transcriptions, languages=dataset["language"])

For example, the following code can be used to evaluate Whisper:

dataset = load_dataset("audioshake/jam-alt", revision="v1.2.0", split="test")
dataset = dataset.cast_column("audio", datasets.Audio(decode=False))  # Get the raw audio file, let Whisper decode it

model = whisper.load_model("tiny")
transcriptions = [
  "\n".join(s["text"].strip() for s in model.transcribe(a["path"])["segments"])
  for a in dataset["audio"]
]
compute_metrics(dataset["text"], transcriptions, languages=dataset["language"])

Alternatively, if you already have transcriptions, you might prefer to skip loading the audio column:

dataset = load_dataset("audioshake/jam-alt", revision="v1.2.0", split="test", columns=["name", "text", "language", "license_type"])

Citation

When using the benchmark, please cite our paper as well as the original JamendoLyrics paper:

@misc{cifka-2024-jam-alt,
  author       = {Ondrej C{\'{\i}}fka and
                  Hendrik Schreiber and
                  Luke Miner and
                  Fabian{-}Robert St{\"{o}}ter},
  title        = {Lyrics Transcription for Humans: {A} Readability-Aware Benchmark},
  booktitle    = {Proceedings of the 25th International Society for 
                  Music Information Retrieval Conference},
  pages        = {737--744},
  year         = 2024,
  publisher    = {ISMIR},
  doi          = {10.5281/ZENODO.14877443},
  url          = {https://doi.org/10.5281/zenodo.14877443}
}
@inproceedings{durand-2023-contrastive,
  author={Durand, Simon and Stoller, Daniel and Ewert, Sebastian},
  booktitle={2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Contrastive Learning-Based Audio to Lyrics Alignment for Multiple Languages}, 
  year={2023},
  pages={1-5},
  address={Rhodes Island, Greece},
  doi={10.1109/ICASSP49357.2023.10096725}
}