espnet3.systems.esp2_cls.system.ESP2CLSSystem
espnet3.systems.esp2_cls.system.ESP2CLSSystem
class espnet3.systems.esp2_cls.system.ESP2CLSSystem(training_config: DictConfig | None = None, inference_config: DictConfig | None = None, metrics_config: DictConfig | None = None, publication_config: DictConfig | None = None, demo_config: DictConfig | None = None)
Bases: BaseSystem
Classification-specific system.
This system adds remove_long_short, which removes utterances outside a duration range, and prepare_labels, which builds the label list consumed as token_list. Both run between create_dataset and collect_stats; their logs go under each stage’s save_path.
Example
s
>>> [s.name for s in ESP2CLSSystem.stages[:4]]
['create_dataset', 'remove_long_short', 'prepare_labels', 'collect_stats']Initialize the system with optional stage configs.
- Parameters:
- training_config – Training configuration for data preparation, statistics collection, and model training.
- inference_config – Inference configuration used by the
inferstage. - metrics_config – Measurement configuration used by the
measurestage. - publication_config – Publication configuration for
pack_modelandupload_modelstages. - demo_config – Demo configuration for the
demostage.
prepare_labels(*args, **kwargs)
Build the label list used as token_list by the classifier.
Labels are collected from the third column of the training manifest, split on whitespace so that multi-label utterances contribute every label they carry. The list is ordered by descending frequency, which matches stage 4 of the ESPnet2 cls.sh recipe.
Under training_config.prepare_labels, configure:
| Field | Description |
|---|---|
save_path | Directory to write the label list into. |
filename | Label list file name, such as token_list. |
manifest_path | Training manifest to read. Defaults to<br/>data/manifest/train.tsv. |
add_symbol | Optional "<symbol>:<index>" entries; a negative index<br/>counts from the end. |
Example
prepare_labels:
save_path: data
filename: token_list
manifest_path: data/manifest/train.tsv
add_symbol:
- "<unk>:-1"- Raises:RuntimeError – If required configuration is missing, the manifest is not found, an
add_symbolentry is malformed, or the manifest holds no label.
remove_long_short(*args, **kwargs)
Remove utterances outside the configured duration range.
This stage reads WAV headers via soundfile (in parallel, through RemoveLongShortProvider / RemoveLongShortRunner) and writes filtered manifests for downstream stages. It mirrors stage 3 of the ESPnet2 cls.sh recipe.
Under training_config.remove_long_short, configure:
| Field | Description |
|---|---|
min_wav_duration | Minimum duration in seconds. |
max_wav_duration | Maximum duration in seconds. |
save_path | Directory in which to save filtered manifests. |
splits | Splits to process. Defaults to [train, valid, test]. |
manifest_paths | Optional mapping from split to manifest path. Defaults to<br/>data/manifest/{split}.tsv. |
Example
remove_long_short:
min_wav_duration: 0.1
max_wav_duration: 30
save_path: data/manifest_filtered
splits: [train, valid, test]- Raises:RuntimeError – If required configuration is missing or a manifest file is not found.
stages : ClassVar[tuple[StageSpec, ...]] = (StageSpec(name='create_dataset', config='training', log_dir='data_dir'), StageSpec(name='remove_long_short', config='training', log_dir='remove_long_short.save_path'), StageSpec(name='prepare_labels', config='training', log_dir='prepare_labels.save_path'), StageSpec(name='collect_stats', config='training', log_dir='stats_dir'), StageSpec(name='train', config='training', log_dir='exp_dir'), StageSpec(name='infer', config='inference', log_dir='inference_dir'), StageSpec(name='measure', config='metrics', log_dir='inference_dir'), StageSpec(name='pack_model', config=('training', 'publication'), log_dir=None), StageSpec(name='upload_model', config='publication', log_dir=None), StageSpec(name='pack_demo', config='demo', log_dir='pack.out_dir'), StageSpec(name='upload_demo', config='demo', log_dir='pack.out_dir'))
