Implementation:Speechbrain Speechbrain Eval ESC50 Interpret

Knowledge Sources	SpeechBrain
Domains	Sound_Classification, Interpretability
Last Updated	2026-02-09 00:00 GMT

Overview

Concrete tool for out-of-distribution (OOD) evaluation of audio classifier interpreters provided by the SpeechBrain library.

Description

This recipe provides evaluation scripts for assessing interpreter models (L2I and LMAC) under out-of-distribution conditions. It defines the ESCContaminated dataset class that creates contaminated versions of the ESC-50 dataset by mixing with LJSpeech, white noise, or other ESC-50 sounds. Also defines the LJSPEECH_split helper class for train/test splitting of LJSpeech data. The evaluation measures how well interpreters can isolate and identify sound class features even when audio is contaminated with out-of-distribution noise sources.

Usage

Use this recipe to evaluate trained L2I or LMAC interpreter models on contaminated ESC-50 data. Requires the ESC-50 dataset and optionally the LJSpeech dataset. Specify the overlap type (mixtures, ljspeech, or white_noise) to control the contamination source.

Code Reference

Source Location

• Repository: SpeechBrain
• File: recipes/ESC50/interpret/eval.py

Signature

class ESCContaminated(torch.utils.data.Dataset):
    def __init__(self, esc50_ds, cont_d, overlap_multiplier=2, overlap_type="mixtures"):
        ...
    def __getitem__(self, idx):
        ...

class LJSPEECH_split(dts.LJSPEECH):
    def __init__(self, root, url, folder_in_archive, download, train=True):
        ...

Import

python recipes/ESC50/interpret/eval.py hparams/l2i_cnn14.yaml --data_folder /path/to/ESC-50-master --overlap_type mixtures --add_wham_noise False

I/O Contract

Inputs

Name	Type	Required	Description
esc50_ds	dataset	Yes	The ESC-50 dataset as prepared for training
cont_d	dataset	Yes	The contamination dataset (ESC-50, LJSpeech, or synthetic)
overlap_type	str	Yes	One of "mixtures", "ljspeech", or "white_noise"

Outputs

Name	Type	Description
contaminated_samples	tuple	Mixed audio with contamination and original labels
interpretation_metrics	dict	Faithfulness and plausibility scores for the interpreter

Usage Examples

python eval.py hparams/l2i_cnn14.yaml --data_folder /path/to/ESC-50-master --overlap_type mixtures --add_wham_noise False

Related Pages

• Principle:Speechbrain_Speechbrain_Audio_Interpretability