espnet3.systems.f5tts.src.remove_long_short.load_manifest_entries

source

espnet3.systems.f5tts.src.remove_long_short.load_manifest_entries(manifest_path: str | Path) → Tuple[List[Tuple[str, str, str]], int]

Parse a TSV manifest into its usable rows.

Each line is expected to be utt_id\twav_path\ttext\tspeaker_id; any columns after the third are carried along untouched. Blank lines are skipped.

• Parameters:manifest_path – Manifest written by the recipe’s create_dataset stage.
• Returns: A tuple (entries, num_dropped_empty). entries is a list of (utt_id, wav_path, line) tuples in manifest order, where line is the original row, newline-terminated for direct write-back. num_dropped_empty counts the rows dropped for having no text.
• Raises:FileNotFoundError – If manifest_path does not exist.

Example

>>> entries, num_dropped_empty = load_manifest_entries(
...     "data/manifest/train.tsv"
... )
>>> [utt_id for utt_id, _, _ in entries], num_dropped_empty
(['utt_a', 'utt_b'], 1)
>>> entries[0]
('utt_a', 'corpus/utt_a.wav',
 'utt_a\tcorpus/utt_a.wav\thello world\tspk1\n')

NOTE

Rows with fewer than three columns or an empty transcript are dropped here, mirroring espnet2’s NF != 1 filter, so they never reach duration filtering and their audio file is never opened.