Implementation:Speechbrain Speechbrain Train ESC50 Classification

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

Overview

Concrete tool for training a sound classifier on the ESC-50 dataset provided by the SpeechBrain library.

Description

This recipe defines the ESC50Brain class (subclass of sb.core.Brain) for environmental sound classification on the ESC-50 dataset. The pipeline computes STFT features, optionally applies mel-spectrogram conversion and log1p normalization, then feeds the result into an embedding model (supporting CNN14, FocalNet, ViT, or Conv2D architectures) followed by a classifier head. Supports data augmentation including waveform augmentation and WHAM! noise addition. Evaluation includes confusion matrix generation.

Usage

Use this recipe to train and evaluate a sound classifier on the ESC-50 dataset. Requires the ESC-50-master data folder. Supports multiple backbone architectures via hyperparameter configuration files (cnn14.yaml, focalnet.yaml, vit.yaml, conv2d.yaml).

Code Reference

Source Location

• Repository: SpeechBrain
• File: recipes/ESC50/classification/train.py

Signature

class ESC50Brain(sb.core.Brain):
    def compute_forward(self, batch, stage):
        ...
    def compute_objectives(self, predictions, batch, stage):
        ...

Import

python recipes/ESC50/classification/train.py hparams/cnn14.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 (class labels)
stage	sb.Stage	Yes	TRAIN, VALID, or TEST

Outputs

Name	Type	Description
predictions	tuple	Classifier output logits and lens
loss	torch.Tensor	Cross-entropy classification loss

Usage Examples

python train.py hparams/cnn14.yaml --data_folder /path/to/ESC-50-master

Related Pages

• Principle:Speechbrain_Speechbrain_Sound_Classification_Training