Implementation:Speechbrain Speechbrain Interpret AMT

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

Overview

Concrete tool for interpreting audio classifiers by-design via activation map thresholding (AMT) provided by the SpeechBrain library.

Description

This recipe defines the InterpreterESC50Brain class (subclass of sb.core.Brain) for interpreting audio classifiers using activation map thresholding. The approach extracts intermediate activation maps (modulators for FocalNet, attention maps for ViT) from a pre-trained classifier, aggregates them to form a saliency mask, and applies it to the input spectrogram to isolate class-relevant regions. The resulting interpretation spectrogram is inverted back to waveform using STFT phase reconstruction. Supports both FocalNet and ViT backbone architectures.

Usage

Use this recipe to generate interpretations of ESC-50 classifier predictions using activation map thresholding. Requires a pre-trained classifier checkpoint and the ESC-50 dataset. Configure with amt_focalnet.yaml or amt_vit.yaml depending on the backbone.

Code Reference

Source Location

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

Signature

class InterpreterESC50Brain(sb.core.Brain):
    def invert_stft_with_phase(self, X_int, X_stft_phase):
        ...
    def preprocess(self, wavs):
        ...
    def classifier_forward(self, X_stft_logpower):
        ...
    def compute_forward(self, batch, stage):
        ...
    def compute_objectives(self, predictions, batch, stage):
        ...

Import

python recipes/ESC50/interpret/interpret_amt.py hparams/amt_focalnet.yaml --data_folder /path/to/ESC-50-master

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, STFT phase, classifier predictions, original spectrogram
loss	torch.Tensor	Interpretation fidelity loss value

Usage Examples

python interpret_amt.py hparams/amt_focalnet.yaml --data_folder /path/to/ESC-50-master

Related Pages

• Principle:Speechbrain_Speechbrain_Audio_Interpretability