Implementation:Speechbrain Speechbrain Train IWSLT22 ST

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

Overview

Concrete tool for speech translation training on the IWSLT22 low-resource task provided by the SpeechBrain library.

Description

This recipe defines the ST class (subclass of sb.core.Brain) for fine-tuning a wav2vec2 model for the speech translation task without transcriptions. The architecture uses wav2vec2 as the speech encoder, a dimensionality reduction layer, and a Transformer decoder-only model for generating target language tokens. Decoding uses SentencePiece tokenization and Moses detokenization. Supports separate valid and test beam searches with BLEU score evaluation.

Usage

Use this recipe to train a direct speech translation model on the IWSLT22 low-resource dataset. Requires pre-trained wav2vec2 weights and a SentencePiece tokenizer. Configure with train_w2v2_st.yaml.

Code Reference

Source Location

• Repository: SpeechBrain
• File: recipes/IWSLT22_lowresource/AST/transformer/train.py

Signature

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

Import

python recipes/IWSLT22_lowresource/AST/transformer/train.py hparams/train_w2v2_st.yaml --data_folder /path/to/data

I/O Contract

Inputs

Name	Type	Required	Description
batch	PaddedBatch	Yes	Batch containing sig (waveforms), tokens_bos, and tokens_eos (target translation tokens)
stage	sb.Stage	Yes	TRAIN, VALID, or TEST

Outputs

Name	Type	Description
predictions	tuple	Log-softmax sequence probabilities, wav_lens, and decoded hypotheses
loss	torch.Tensor	Sequence-level NLL loss on target translation tokens

Usage Examples

python train.py hparams/train_w2v2_st.yaml --data_folder /path/to/data

Related Pages

• Principle:Speechbrain_Speechbrain_Speech_Translation_Training