Implementation:Speechbrain Speechbrain Train Switchboard Seq2Seq

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

Overview

Concrete tool for training a sequence-to-sequence ASR model on the Switchboard dataset provided by the SpeechBrain library.

Description

This recipe defines the ASR class (subclass of sb.Brain) for attention-based sequence-to-sequence speech recognition on Switchboard conversational telephone speech. The encoder is a CRDNN model and the decoder is a standard GRU with beam search. The model is jointly trained with CTC and NLL losses, with CTC used only during early training epochs. The class accepts a custom normalize_fn for text normalization of Switchboard transcripts. BPE subword units are used as recognition tokens.

Usage

Use this recipe to train a seq2seq ASR model with CTC+attention on the Switchboard dataset (~300 hours). Requires the corresponding hyperparameter YAML file, a pre-trained tokenizer, and data preparation script.

Code Reference

Source Location

• Repository: SpeechBrain
• File: recipes/Switchboard/ASR/seq2seq/train.py

Signature

class ASR(sb.Brain):
    def __init__(self, modules=None, opt_class=None, hparams=None,
                 run_opts=None, checkpointer=None, normalize_fn=None):
        ...
    def compute_forward(self, batch, stage):
        """Forward computations from the waveform batches to the output probabilities."""
        ...
    def compute_objectives(self, predictions, batch, stage):
        """Computes the loss (CTC+NLL) given predictions and targets."""
        ...

Import

# Run as recipe script
python recipes/Switchboard/ASR/seq2seq/train.py hparams/train_BPE_2000.yaml --data_folder /path/to/switchboard

I/O Contract

Inputs

Name	Type	Required	Description
batch.sig	torch.Tensor	Yes	Input waveform signal
batch.tokens_bos	torch.Tensor	Yes	Target token sequence with BOS prefix
batch.tokens_eos	torch.Tensor	Yes	Target token sequence with EOS suffix
batch.tokens	torch.Tensor	Yes	Target token sequence (for CTC)

Outputs

Name	Type	Description
p_ctc	torch.Tensor	CTC log-probabilities (during early training epochs)
p_seq	torch.Tensor	Seq2seq log-probabilities from attention decoder
wav_lens	torch.Tensor	Relative waveform lengths
p_tokens	list	Beam search hypotheses (at validation/test)

Usage Examples

python train.py hparams/train_BPE_2000.yaml --data_folder /path/to/Switchboard

Related Pages

• Principle:Speechbrain_Speechbrain_Seq2Seq_ASR_Training