Larry VCTK Normalized

Built from the CSTR VCTK Corpus, version 0.92, published by the Centre for Speech Technology Research (CSTR), University of Edinburgh.

The VCTK dataset from Edinburgh datashare redistributed on huggingface with no preprocessing. 109 English speakers with various accents. Both of the corpus's microphone recordings are included as separate rows, giving 88,156 clips (~82.5 hours) of 44,283 utterances, un-pruned.

Fields

• id: utterance id, e.g. p225_001; the same utterance appears once per microphone, so id and mic_id together identify a row
• speaker_id: VCTK speaker id, e.g. p225; speaker age, gender, and accent are in the corpus's speaker-info.txt
• mic_id: mic1 (DPA 4035 omni-directional) or mic2 (Sennheiser MKH 800)
• path: the audio file's path within the corpus's wav48_silence_trimmed folder
• audio: Arrow audio typed column of (bytes, path); 48 kHz mono FLAC as distributed in the corpus
• text_original: the transcript as distributed in the corpus
• text_normalized: the same text spelled out as spoken, in uppercase letters, apostrophes, and spaces (e.g. "Mr" -> MISTER)
• duration_s: clip length in seconds
• ctc_mismatch: how poorly the clip's audio matches its text (lower is better)

Construction

Each transcript in txt/ is paired with its recordings in wav48_silence_trimmed/ by file name. Scored with facebook/wav2vec2-large-960h-lv60-self. See provenance.json for the producing commit, timestamp, and ancestry.

Known corrections

From the corpus

• Silent, cut-off, and misspoken recordings were deleted, and several transcripts fixed, by the corpus authors; see the corpus's update.txt
• The audio was end-pointed (leading and trailing silence trimmed) by the corpus authors

From this dataset

• Transcripts per example are joined onto a single line with single spaces

Known gaps

• p315 is not included: the corpus has their recordings but no transcripts
• 300 transcripts for p362 are not included: the corpus has no recordings for them
• p280 has no mic2 recordings (the corpus notes technical issues with that microphone for this speaker), so their 410 utterances appear once instead of twice
• mic1 recordings have very low frequency noise, per the corpus's update.txt
• The Rainbow Passage and elicitation paragraph are read by every speaker, so those sentences repeat across speakers; there are 13,754 distinct normalized sentences
• No quality, duration, or ctc_mismatch filtering is applied here; that is left to downstream consumers

The CSTR VCTK Corpus is licensed CC-BY-4.0, and this redistribution is CC-BY-4.0 under the larry project.

Citation

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

@misc{veaux2019vctk,
  author    = {Veaux, Christophe and Yamagishi, Junichi and MacDonald, Kirsten},
  title     = {{CSTR VCTK Corpus}: English Multi-speaker Corpus for {CSTR} Voice Cloning Toolkit (version 0.92)},
  publisher = {The Centre for Speech Technology Research (CSTR), University of Edinburgh},
  year      = {2019},
  url       = {https://datashare.ed.ac.uk/handle/10283/3443}
}

Preprocessing pipeline code is at https://github.com/Mekadrom/larry, and this dataset is fully reproducible with a single command in that repo, given the corpus archive DS_10283_3443.zip downloaded from the link above into the configured download_extract_cache_dir (the downloader may attempt to download it for you, but this has not been tested):

uv run preprocessing --config_file configs/preprocessing/normalized/voice/edinburgh-datashare_vctk-normalized.yaml --output_dir ~/.cache/larry/datasets/normalized/voice/edinburgh-datashare_vctk-normalized --clean