espnet3.systems.f5tts.src.remove_long_short.remove_long_short

source

espnet3.systems.f5tts.src.remove_long_short.remove_long_short(config: DictConfig) → None

Write duration-filtered copies of the recipe’s manifests.

The body of the remove_long_short stage. For every configured split it reads the manifest written by create_dataset, drops rows without text, checks each remaining utterance’s duration from its audio header (in parallel, honouring config.parallel), and writes the surviving rows, unchanged and in order, to save_path/<manifest file name>.

Configuration should include (under remove_long_short):

• min_wav_duration: Minimum duration in seconds. Utterances at or below it are dropped.
• max_wav_duration: Maximum duration in seconds. Utterances at or above it are dropped.
• save_path: Directory in which to save the filtered manifests.
• splits: Splits to process, a name or a list of names. Defaults to [train, valid, test].
• manifest_paths: Optional mapping from split to manifest path. A split without an entry uses data/manifest/{split}.tsv, relative to the working directory.
• batch_size: Optional number of utterances per runner call.

• Parameters:config – Training config holding the remove_long_short block and, optionally, parallel.
• Raises:RuntimeError – If the remove_long_short block, save_path or a duration bound is missing, or a split’s manifest does not exist.

Example

remove_long_short:
  min_wav_duration: 1.0
  max_wav_duration: 20.0
  save_path: ${data_dir}/manifest_filtered
  splits: [train, valid, test]
  manifest_paths:
    train: ${data_dir}/manifest/train.tsv

>>> remove_long_short(training_config)

NOTE

Running the stage again overwrites the filtered manifests. Shard results from an earlier run are never reused, because the keep/drop decisions depend on the duration bounds.