Implementation:Speechbrain Speechbrain Train LMAC

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

Overview

Concrete tool for training the L-MAC interpreter model for audio classifiers provided by the SpeechBrain library.

Description

This recipe defines the LMAC class (subclass of InterpreterBrain) for training the L-MAC (Listenable Maps for Audio Classifiers) interpreter. L-MAC generates interpretation masks that isolate class-relevant regions in audio spectrograms. It includes a cross-correlation metric (supporting both convolution-based and dot-product variants) for fine-tuning, total variation (TV) loss for mask smoothness, and support for NMF-based spectrogram decomposition. The interpretation pipeline produces listenable audio excerpts highlighting the most salient regions for each class prediction.

Usage

Use this recipe to train an L-MAC interpreter on the ESC-50 dataset. Requires the ESC-50 data folder and a pre-trained classifier checkpoint. Supports WHAM! noise augmentation as used in the L-MAC paper. Configure with lmac_cnn14.yaml.

Code Reference

Source Location

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

Signature

class LMAC(InterpreterBrain):
    def crosscor(self, spectrogram, template):
        ...
    def interpret_computation_steps(self, wavs_batch):
        ...
    def compute_forward(self, batch, stage):
        ...
    def compute_objectives(self, predictions, batch, stage):
        ...

Import

python recipes/ESC50/interpret/train_lmac.py hparams/lmac_cnn14.yaml --data_folder /path/to/ESC50 --add_wham_noise True --wham_folder /path/to/wham_noise

I/O Contract

Inputs

Name	Type	Required	Description
batch	PaddedBatch	Yes	Batch containing sig (waveforms) and class_string_encoded (labels)
stage	sb.Stage	Yes	TRAIN, VALID, or TEST

Outputs

Name	Type	Description
predictions	tuple	Interpretation spectrogram, mask, STFT phase, classifier predictions
loss	torch.Tensor	Combined interpretation fidelity, TV loss, and classification loss

Usage Examples

python train_lmac.py hparams/lmac_cnn14.yaml --data_folder /path/to/ESC50 --add_wham_noise True --wham_folder /path/to/wham_noise

Related Pages

• Principle:Speechbrain_Speechbrain_Audio_Interpretability