You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access is reviewed manually. The audio comes from commercial audiobooks and remains the property of its rights holders. It is provided only for non-commercial research on expressive speech. Do not redistribute it.

Log in or Sign Up to review the conditions and access this dataset content.

A-Bursts Labeled

A-Bursts with a transcript, a free-form delivery description and a short voice-acting direction for every clip, plus a balancing set of calm, warm speech. It is meant for directed or instruct-style expressive TTS ("say this line, delivered like this").

config clips hours what
main 1,663,715 4,686 emotional bursts (shouting, crying, pleading, laughing, tender…)
gentle 115,463 ~360 calm, warm, positive speech (high valence, low arousal) to balance main
default both main + gentle
refs calm narration reference clips per book/narrator, for pairing bursts with a voice prompt

Columns

All A-Bursts columns (audio, id, book, book_id, start, end, duration, emotion a/d/v, quality checks, emo_windows, …), plus:

column description
subset main or gentle
text transcript (nvidia/parakeet-tdt-0.6b-v3)
delivery long free-form description of how the line is performed: emotion, pace, pitch, dynamics, non-verbal sounds (nvidia/audio-flamingo-next-captioner-hf, non-commercial)
delivery_truncated true if the captioner hit its length limit
style a few words, e.g. bitter, dry sarcasm, soft, consoling
direction one or two sentences of actor direction written from delivery, e.g. "Quiet and reflective at first, then gently weightier; deep voice, deliberate pauses."

style and direction were rewritten from delivery by Claude Haiku. Where it declined (about 0.02% of rows, lines whose content concerned weapons or violence), they were written by hand from the same notes.

How gentle was selected

From the same emotion scores as A-Bursts: windows with valence ≥ 1.5 sd above the median, arousal ≤ 0.5 sd below it, emotion distance < 3, at least 90% speech and good quality, at most 20 per book. Clips were cut the same way as main (VAD chunk, speaker trim, pause snap, packet copy).

Notes

  • Transcripts and descriptions are model output and contain errors. The delivery captioner sometimes invents details (speaker gender, background sounds); direction inherits them.
  • See the A-Bursts card for selection details, audio format and clip-edge precision.
  • Rows in main and gentle are each shuffled.
Downloads last month
593