README.md

---
task_categories:
- automatic-speech-recognition
multilinguality:
- multilingual
language:
- en
- fr
- de
- es
tags:
- music
- lyrics
- evaluation
- benchmark
- transcription
- pnc
pretty_name: 'Jam-ALT: A Readability-Aware Lyrics Transcription Benchmark'
paperswithcode_id: jam-alt
configs:
- config_name: all
  data_files:
  - split: test
    path:
      - metadata.jsonl
      - subsets/*/audio/*.mp3
  default: true
- config_name: de
  data_files:
  - split: test
    path:
      - subsets/de/metadata.jsonl
      - subsets/de/audio/*.mp3
- config_name: en
  data_files:
  - split: test
    path:
      - subsets/en/metadata.jsonl
      - subsets/en/audio/*.mp3
- config_name: es
  data_files:
  - split: test
    path:
      - subsets/es/metadata.jsonl
      - subsets/es/audio/*.mp3
- config_name: fr
  data_files:
  - split: test
    path:
      - subsets/fr/metadata.jsonl
      - subsets/fr/audio/*.mp3
---

# Jam-ALT: A Readability-Aware Lyrics Transcription Benchmark


## Dataset description

* **Project page:** https://audioshake.github.io/jam-alt/
* **Source code:** https://github.com/audioshake/alt-eval
* **Paper (ISMIR 2024):** https://doi.org/10.5281/zenodo.14877443
* **Paper (arXiv):** https://arxiv.org/abs/2408.06370
* **Extended abstract (ISMIR 2023 LBD):** https://arxiv.org/abs/2311.13987
* **Related datasets:** https://huggingface.co/jamendolyrics

Jam-ALT is a revision of the [**JamendoLyrics**](https://huggingface.co/datasets/jamendolyrics/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](#citation)): \
📄 [**Lyrics Transcription for Humans: A Readability-Aware Benchmark**](https://doi.org/10.5281/zenodo.14877443) \
👥 O. Cífka, H. Schreiber, L. Miner, F.-R. Stöter \
🏢 [AudioShake](https://www.audioshake.ai/)

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

💥 **New:** The dataset now has **line-level timings**.
They were added in the paper 📄 **Exploiting Music Source Separation for Automatic Lyrics Transcription with Whisper** by
J. Syed, I. Meresman-Higgs, O. Cífka, and M. Sandler, presented at the [2025 ICME Workshop AI for Music](https://ai4musicians.org/2025icme.html).

> [!note]
> **Note:** The dataset is not time-aligned at the word level. To evaluate **automatic lyrics alignment** (**ALA**), please use [JamendoLyrics](https://huggingface.co/datasets/jamendolyrics/jamendolyrics), which is the standard benchmark for that task.

> [!tip]
> See the [project website](https://audioshake.github.io/jam-alt/) for details and the [JamendoLyrics community](https://huggingface.co/jamendolyrics) for related datasets.

## Loading the data

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

A subset is defined for each language (`en`, `fr`, `de`, `es`);
for example, use `load_dataset("jamendolyrics/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.

## Running the benchmark

The evaluation is implemented in our [`alt-eval` package](https://github.com/audioshake/alt-eval):
```python
from datasets import load_dataset
from alt_eval import compute_metrics

dataset = load_dataset("jamendolyrics/jam-alt", revision="v1.3.0", split="test")
# transcriptions: list[str]
compute_metrics(dataset["text"], transcriptions, languages=dataset["language"])
```

For example, the following code can be used to evaluate Whisper:
```python
dataset = load_dataset("jamendolyrics/jam-alt", revision="v1.3.0", 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:
```python
dataset = load_dataset("jamendolyrics/jam-alt", revision="v1.3.0", split="test").remove_columns("audio")
```

## Using non-lexical tags

In addition to the `text` field, each song and line timing also contains a field `text_tagged` with
non-lexical vocables in the text enclosed in inline opening and closing tags `<nl>` and `</nl>`,
as in shown in this verse from `Les_files_d'attente_-_Law'`:

```
<nl> (Ouh) </nl> chacun est pour l'autre le suivant
Personne ne bouge, tout l'monde attend <nl> (ouh-ouh-ouh-ouh) </nl>
Marchez-vous dessus, montrez les dents
Crachez-vous dessus, l'est encore temps
```

## Citation

When using the benchmark, please cite our [ISMIR paper](https://doi.org/10.5281/zenodo.14877443) as well as the original [JamendoLyrics paper](https://arxiv.org/abs/2306.07744).
For the line-level timings or non-lexical tags, please cite the ICME workshop paper.
```bibtex
@inproceedings{cifka-2024-jam-alt,
  author       = {Ond{\v{r}}ej 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}
}
@inproceedings{syed-2025-mss-alt,
  author       = {Jaza Syed and
                  Ivan Meresman-Higgs and
                  Ond{\v{r}}ej C{\'{\i}}fka and
                  Mark Sandler},
  title        = {Exploiting Music Source Separation for Automatic Lyrics Transcription with {Whisper}},
  booktitle    = {2025 {IEEE} International Conference on Multimedia and Expo Workshops (ICMEW)},
  publisher    = {IEEE},
  year         = {2025},
  note         = {to appear}
}
```

## Contributions

The transcripts, originally from the [JamendoLyrics](https://huggingface.co/datasets/jamendolyrics/jamendolyrics) dataset, were revised by Ondřej Cífka, Hendrik Schreiber, Fabian-Robert Stöter, Luke Miner, Laura Ibáñez, Pamela Ode, Mathieu Fontaine, Claudia Faller, April Anderson, Constantinos Dimitriou, and Kateřina Apolínová.
Line-level timings were automatically transferred from JamendoLyrics and manually corrected by Ondřej Cífka and Jaza Syed to fit the revised transcripts.