thelarryproject/vctk-normalized-asr

thelarryproject VCTK Normalized for ASR Training

Built from thelarryproject/vctk, the unfiltered CSTR VCTK Corpus (version 0.92).

The corpus's mic1 recordings, filtered for ASR training, with 5 speakers held out for validation:

split	clips	hours	speakers
train	41,978	39.3	104
validation	2,105	2.0	5

The validation speakers are p238, p253, p275, p277, and p280.

Fields

• id: utterance id, e.g. p225_001
• speaker_id: VCTK speaker id, e.g. p225; speaker age, gender, and accent are in the corpus's speaker-info.txt
• mic_id: always mic1 (DPA 4035 omni-directional)
• 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, scored with facebook/wav2vec2-large-960h-lv60-self (lower is better)

Construction

Only mic1 rows are kept, which makes id unique. Removed:

• clips shorter than 0.1 s or longer than 25 s
• clips failing an audio quality check: overall level below -45 dBFS, speech in under 30% of frames, signal-to-noise ratio below 5 dB, or more than 0.1% of samples clipped
• clips with ctc_mismatch above 0.5

Each speaker then goes to validation if a hash of their id falls in the lowest 5 of 100 buckets, so the split is fixed and speaker-disjoint. 200 of the 44,283 mic1 clips are removed in total. 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
• 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 and across the train and validation splits

The CSTR VCTK Corpus is licensed CC-BY-4.0 by the University of Edinburgh, and this redistribution is CC-BY-4.0 under thelarryproject.

Citation

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

@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:

uv run preprocessing -f configs/preprocessing/normalized/voice/asr/thelarryproject_vctk-normalized-asr.yaml --output_dir ~/.cache/larry/datasets/normalized/voice/asr/thelarryproject_vctk-normalized-asr --clean