Implementation:Speechbrain Speechbrain Train L2I

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

Overview

Concrete tool for training the L2I (Listen to Interpret) model to interpret audio classifiers provided by the SpeechBrain library.

Description

This recipe defines the L2I class (subclass of InterpreterBrain) for training the L2I interpreter as described in the paper (https://arxiv.org/abs/2202.11479). The L2I model learns to generate interpretation spectrograms that highlight class-relevant regions in audio. It uses NMF (Non-negative Matrix Factorization) with a learned psi network to produce activation maps, which are combined with NMF dictionary components to form interpretation masks. The interpretation is applied to the original spectrogram and inverted back to audio via STFT phase reconstruction.

Usage

Use this recipe to train an L2I interpreter on the ESC-50 dataset. Requires the ESC-50 data folder and a pre-trained classifier checkpoint. Optionally supports WHAM! noise augmentation. Configure with l2i_cnn14.yaml or l2i_conv2d.yaml.

Code Reference

Source Location

• Repository: SpeechBrain
• File: recipes/ESC50/interpret/train_l2i.py

Signature

class L2I(InterpreterBrain):
    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_l2i.py hparams/l2i_cnn14.yaml --data_folder /path/to/ESC50

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, NMF mask, STFT phase, classifier predictions
loss	torch.Tensor	Combined interpretation fidelity and classification loss

Usage Examples

python train_l2i.py hparams/l2i_cnn14.yaml --data_folder /path/to/ESC50

Related Pages

• Principle:Speechbrain_Speechbrain_Audio_Interpretability