alt-eval

alt-eval

An automatic lyrics transcription (ALT) evaluation toolkit, released with the Jam-ALT benchmark.

The package implements metrics designed to work well with lyrics formatted according to music industry standards (see the Jam-ALT annotation guide), namely:

• A word error rate (WER) computed on text tokenized in a way that accounts for non-standard spellings common in song lyrics.
• A case-sensitive WER.
• Precision, recall and F-score for symbols important for written lyrics:
  • Punctuation
  • Parentheses (used to delimit background vocals)
  • Line breaks
  • Section breaks (i.e. double line breaks)

Under the hood, the text is pre-processed using the sacremoses tokenizer and punctuation normalizer. Note that apostrophes and single quotes are never treated as quotation marks, but as part of a word, marking an elision or a contraction.

Usage

Install the package with pip install alt-eval.

To compute the metrics:

from alt_eval import compute_metrics
compute_metrics(references, hypotheses)

where references and hypotheses are lists of strings. To specify the language (English by default), use the languages parameter, passing either a single language code, or a list of language codes corresponding to individual examples.

For Jam-ALT, use:

from datasets import load_dataset
dataset = load_dataset("audioshake/jam-alt")["test"]
compute_metrics(dataset["text"], transcriptions, languages=dataset["language"])

If you are only interested in WER, you may skip formatting- and punctuation-related metrics by passing include_other=False.

Use visualize_errors=True to also get a list of HTML snippets that can be used to visualize the errors in each transcript.

Language support

The package implements language-specific tokenization via sacremoses, enhanced with custom rules. Support is well tested for English, Spanish, German, and French.

For writing systems that do not use spaces to separate words (Chinese, Japanese, Thai, Lao, Burmese, …), each character is considered as a separate word, as per Radford et al. (2022), making the WER equivalent to CER (character error rate).

See the test cases for examples of how different languages are tokenized. Contributions adding support for additional languages are welcome.

Optional lyrics normalization

The Jam-ALT annotation guide forbids certain end-of-line punctuation and requires the first letter of each line to be uppercase. For transcription systems that do not respect these rules, the results on Jam-ALT can be improved by normalizing the transcripts using the normalize_lyrics() function, which fixes these specific issues. Note, however, that this relies on the line break predictions being correct. Moreover, other datasets may follow different rules. For these reasons, this normalization is not included as a fixed pre-processing step in compute_metrics(), and instead made optional.