aces-metric 0.0.5

ACES metric for evaluating automated audio captioning models based on the semantics of sounds

ACES

This is the repository of Audio Captioning Evaluation on Semantics of Sound (ACES).

Corresponding paper: https://arxiv.org/abs/2403.18572.

In here you will find the instructions how to train an ACES model and calculate statistics.

Installation

pip install aces-metric

Usage

The candidates can be a list, the references can be a list or a list of lists.

from aces import get_aces_score
candidates = ["a bunch of birds are singing"]
references = ["birds are chirping and singing loudly in the forest"]
score = get_aces_score(candidates, references, average=True)

Semantics of sounds

To get an output of classes of semantic groups from a caption:

from transformers import pipeline
pipe = pipeline("token-classification", "gijs/aces-roberta-13", aggregation_strategy="simple")
pipe("Bird chirps in the tree while a car hums")

Evaluation

All the code that is used to evaluate different models for the research paper can be found in the evaluation folder on the github. Particularly, the model evaluation can be found in evaluation/eval.py, and information about the FENSE experiment can be found in evaluation/fense_experiment/main.py.