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
A-Bursts contains short bursts of emotionally expressive, single-speaker speech cut from about 31k English audiobooks. Examples include shouting, crying, pleading and laughing, as well as calm or tender delivery. Every clip comes with emotion and audio-quality annotations.
Totals: 1,663,715 clips, 4,686 hours, from 30,919 books (95 GB). Clip length median about 8 s, mean 10 s; merged bursts can run up to about 60 s.
How clips were selected
- Speech detection. Silero VAD ran over every book.
- Emotion scoring. One 4 s speech window per 30 s of audio was scored with audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim for arousal, dominance and valence. Deterministic quality checks also ran on each window: SNR, level, bandwidth, clipping and gating.
- Selection. A window is kept when all of these hold:
- its emotion distance
distis ≥ 3.0.distis the Euclidean distance from the median (a, d, v) over all windows, with each axis divided by its standard deviation. The median represents flat narration, and 3.0 is roughly the top 6% of speech. - it is at least 90% speech;
- its quality verdict is
good.
- its emotion distance
- Burst cut. Each kept window is expanded to its VAD speech chunk (pauses under 0.7 s are bridged, up to 30 s). The chunk is then trimmed with ECAPA speaker embeddings to the run where the voice matches the selected window, which removes narration around a character's line. Both ends are then snapped to the nearest pause. Overlapping bursts are merged.
- Audio. Rows are globally shuffled. Clips are Ogg Opus, cut from the source stream by copying packets. Nothing is re-encoded, so the audio is bit-identical to the source. The source bitrate is about 48 kbps mono or 96 kbps stereo, at 48 kHz.
Columns
| column | description |
|---|---|
audio |
Ogg Opus clip |
id |
<book_id>_<start in centiseconds> |
book, book_id |
audiobook title and a stable hash id |
start, end, duration |
clip position in the book, in seconds |
channels, sample_rate |
1 or 2; always 48000 |
dist, arousal, dominance, valence |
emotion of the window that triggered the burst (a/d/v roughly 0–1) |
speaker_sim |
mean ECAPA cosine similarity of the burst to that window |
emo_window_start |
start of the triggering 4 s window, in book seconds |
quality, snr_db, level_dbfs, bandwidth_hz, clip_frac |
quality checks on the triggering window |
emo_windows |
every scored 4 s window overlapping the clip, each with start, a/d/v, dist, quality, snr_db and bandwidth_hz |
Notes
distmeasures distance in any direction. Very soft, sad or tender delivery also scores high, not only loud outbursts. For high-energy bursts, filter onarousal > 0.48; for soothing speech, filter onarousal < 0.48.- A score of 3–4 sometimes catches narrators reading in a heightened style. Scores of 4 and above are mostly true outbursts.
- Clip edges have about 20 ms of precision because cuts fall on Opus packet boundaries. The first ~60 ms is lead-in inside a pause.
- The speaker trim can end early on long rants when the voice shifts strongly.
- No transcripts are included.
- Downloads last month
- 804