Implementation:Speechbrain Speechbrain InterpreterBrain

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

Overview

Concrete tool providing the parent class for interpretability recipes in the SpeechBrain library.

Description

This module defines the InterpreterBrain class (subclass of sb.core.Brain) which serves as the shared parent class for all interpretability training recipes (L2I and LMAC). It provides common functionality including STFT inversion with phase reconstruction, audio preprocessing (STFT computation, mel-spectrogram, log-power features), classifier forward pass abstraction supporting multiple backbone architectures (CNN14, FocalNet, ViT, Conv2D), and quantitative evaluation using the Quantus library for faithfulness metrics. The class handles both standard spectral features and mel-spectrogram pathways.

Usage

This class is not intended to be used directly. Instead, it is subclassed by L2I and LMAC interpreter training recipes which implement the specific interpretation computation logic. It provides the shared infrastructure for preprocessing, classifier inference, and evaluation.

Code Reference

Source Location

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

Signature

class InterpreterBrain(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

from interpreter_brain import InterpreterBrain

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, phase, classifier outputs, and original features
loss	torch.Tensor	Interpretation fidelity loss value

Usage Examples

# Subclass InterpreterBrain for a custom interpreter:
from interpreter_brain import InterpreterBrain

class MyInterpreter(InterpreterBrain):
    def interpret_computation_steps(self, wavs_batch):
        # Custom interpretation logic
        ...

Related Pages

• Principle:Speechbrain_Speechbrain_Audio_Interpretability