The dataset viewer is not available for this split.
Error code: RowsPostProcessingError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
SKTTS English features
Training features for the English text-to-speech models of the SKTTS project: log-mel spectrograms and phoneme sequences of two clean, openly licensed English speech corpora. It contains no audio, only the mel features (which a vocoder such as BigVGAN-v2 turns back into approximate audio).
455,183 clips, 525 h train + 2.5 h validation, 1,157 speakers.
Sources and attribution (CC BY 4.0)
- LibriTTS-R (
train-clean-100,train-clean-360): Koizumi et al., "LibriTTS-R: A Restored Multi-Speaker Text-to-Speech Corpus", Interspeech 2023. https://huggingface.co/datasets/mythicinfinity/libritts_r - HiFi-TTS (
train.clean,train.other): Bakhturina et al., "Hi-Fi Multi-Speaker English TTS Dataset", Interspeech 2021. https://huggingface.co/datasets/MikhailT/hifi-tts
Both are CC BY 4.0, and so is this derived dataset: use it freely, with attribution to the two corpora above.
Speaker ids are the sources' own (en_<id>, HiFi-TTS readers as en_hifi_<id>_clean|other).
Layout
shards/{train,validation}-<part>.tar: per clip<id>.npz(mel, float16, shape[100, frames]) and<id>.json(id,speaker,source,split,text,phonesin ARPAbet with stress,n_frames).index.jsonl: one row per clip (id,shard,split,speaker,source,n_frames,n_tokens); one clip in twelve also hassisdrandpesq, no-reference quality estimates (torchaudio SQUIM) used to rank the speakers.features_stats.json: counts and hours.
How the features were made
Audio resampled to 24 kHz, brought to a common loudness (-24 dB RMS over the louder half of the 50 ms frames, peak
limited to 0.95), then the BigVGAN-v2 bigvgan_v2_24khz_100band_256x mel (n_fft 1024, hop 256, 100 bands, up to
12 kHz, natural log). Clips shorter than 1 s or longer than 25.6 s are dropped. 0.5 % of the clips (by id hash) form
the validation split. Text is the sources' normalised text; phones come from g2p_en after the cleanup in
sktts/frontend/english.py (pause marks kept as commas). Code: the SKTTS repository, sktts/data/english.py.
- Downloads last month
- 108