alt-eval 1.0.0

Automatic lyrics transcription evaluation toolkit

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 error rate, measuring the rate of incorrectly predicted letter case.
• 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)

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 JamALT, use:

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

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