thelarryproject/libriheavy-large-index-vocoder

thelarryproject Libriheavy (large) Vocoder Index

An index over the large config of mythicinfinity/libriheavy at revision 784858eb1793bfcf2fd859f7a7ce2dae67dfe5f7: the clips that pass the audio quality and transcript filters used for vocoder training, with their normalized transcripts and durations but no audio. Each row points at its clip in the source dataset by row number.

9,325,776 of the source's 11,156,740 clips remain:

split	clips	hours	speakers
train	9,325,776	34,572.8	6,099

Fields

• row: the clip's row number in the source's train split at the revision above
• id: Libriheavy's clip id, e.g. large/10018/conquestofcanaan_1710_librivox_64kb_mp3/conquestofcanaan_01_tarkington_64kb_5; check it against the source row to confirm the match
• text_normalized: the source's text_original spelled out as spoken, in uppercase letters, apostrophes, and spaces (e.g. "Mr" -> MISTER)
• duration_s: clip length in seconds, as given by the source

Construction

Rows are numbered first. The source's text_transcription column is dropped and audio_duration is renamed to duration_s. Removed, with the stricter thresholds used for synthesis data:

• clips shorter than 0.1 s or longer than 25 s
• clips failing an audio quality check: overall level below -40 dBFS, speech in under 50% of frames, signal-to-noise ratio below 15 dB, or more than 0.01% of samples clipped
• clips with ctc_mismatch above 0.2, scored with facebook/wav2vec2-large-960h-lv60-self
• clips whose audio could not be decoded

Then every column but row, id, duration_s, and text_normalized is dropped; the source has no equivalent of the last two, so they stay here. See provenance.json for the producing commit and timestamp.

Known corrections

N/a

Known gaps

• Libriheavy's own test sets are not speaker-disjoint from this data: 30 of the 72 test_clean_large speakers and 25 of the 73 test_other_large speakers also appear here. LibriSpeech's dev-clean and dev-other speakers do not

Libriheavy is licensed Apache-2.0, and this redistribution is Apache-2.0 under thelarryproject. The underlying audio is public domain LibriVox recordings.

Citation

If you use this dataset, please cite the original corpus as well:

@misc{kang2023libriheavy,
  title         = {Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context},
  author        = {Wei Kang and Xiaoyu Yang and Zengwei Yao and Fangjun Kuang and Yifan Yang and Liyong Guo and Long Lin and Daniel Povey},
  year          = {2023},
  eprint        = {2309.08105},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS}
}

Preprocessing pipeline code is at https://github.com/Mekadrom/larry, and this dataset is fully reproducible with a single command in that repo:

uv run preprocessing -f configs/preprocessing/index/voice/asr/mythicinfinity_libriheavy-large-index-asr.yaml -f configs/preprocessing/index/voice/vocoder/mythicinfinity_libriheavy-large-index-vocoder.yaml --output_dir ~/.cache/larry/datasets/index/voice/vocoder/mythicinfinity_libriheavy-large-index-vocoder --clean