Implementation:Speechbrain Speechbrain Hparams Switchboard Seq2Seq

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

Overview

Hyperparameter configuration for CRDNN seq2seq ASR training on the Switchboard dataset.

Description

HyperPyYAML configuration file that defines the model architecture, training schedule, and data processing pipeline for attention-based seq2seq ASR with a CRDNN encoder and GRU decoder on Switchboard conversational telephone speech. The model uses CTC + NLL losses with BPE unigram tokenization (2000 tokens). A pre-trained tokenizer must be provided. The configuration supports noise augmentation, dynamic batching, and limits duplicate utterances to a maximum of 300 occurrences. Test evaluation is split into Switchboard, Callhome, and full test sets.

Usage

Pass this YAML file as the first argument to the corresponding training script.

Code Reference

Source Location

• Repository: SpeechBrain
• File: recipes/Switchboard/ASR/seq2seq/hparams/train_BPE_2000.yaml

Key Parameters

seed: 1312
number_of_epochs: 20
number_of_ctc_epochs: 5
batch_size: 10
lr: 1.0
ctc_weight: 0.5
sorting: ascending
normalize_words: True
max_utt: 300

# Dynamic batching parameters
feats_hop_size: 0.01
max_batch_length: 20000
num_buckets: 20

I/O Contract

Inputs

Name	Type	Required	Description
--data_folder	str	Yes	Path to Switchboard dataset
--pretrained_tokenizer_path	str	Yes	Path to pre-trained BPE tokenizer

Outputs

Name	Type	Description
Instantiated objects	Python objects	Model, optimizer, scheduler, etc.

Usage Examples

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

Related Pages

• Principle:Speechbrain_Speechbrain_HyperPyYAML_Configuration