Implementation:Speechbrain Speechbrain Train GSC

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

Overview

Concrete tool for training a spoken command classifier on the Google Speech Commands v0.02 dataset provided by the SpeechBrain library.

Description

This recipe defines the SpeakerBrain class (subclass of sb.core.Brain) for keyword/command classification using the Google Speech Commands dataset. The pipeline computes features (supporting LEAF or traditional spectral features), normalizes them, extracts embeddings through an embedding model (supporting xvector or ECAPA-TDNN architectures), and passes them through a classifier. Waveform augmentation is applied during training. The model outputs command class probabilities.

Usage

Use this recipe to train a spoken command recognition classifier on the Google Speech Commands v0.02 dataset. Requires the corresponding hyperparameter YAML file (e.g., xvect.yaml).

Code Reference

Source Location

• Repository: SpeechBrain
• File: recipes/Google-speech-commands/train.py

Signature

class SpeakerBrain(sb.core.Brain):
    def compute_forward(self, batch, stage):
        """Computation pipeline based on a encoder + command classifier."""
        ...
    def compute_objectives(self, predictions, batch, stage):
        """Computes the loss using command-id as label."""
        ...

Import

# Run as recipe script
python recipes/Google-speech-commands/train.py hparams/xvect.yaml --data_folder /path/to/GSC

I/O Contract

Inputs

Name	Type	Required	Description
batch.sig	torch.Tensor	Yes	Input waveform signal
batch.command_encoded	torch.Tensor	Yes	Encoded command label

Outputs

Name	Type	Description
outputs	torch.Tensor	Class posterior probabilities over speech commands
lens	torch.Tensor	Relative signal lengths

Usage Examples

python train.py hparams/xvect.yaml --data_folder /path/to/Google-speech-commands

Related Pages

• Principle:Speechbrain_Speechbrain_Sound_Classification_Training