emocean

Emotionally expressive English TTS dataset with speaker IDs, prosody features, and emotion labels.

Dataset Summary

Metric	Value
Total segments	2,261
Total duration	3.39 hours
Speakers	13
Sources	12 videos
Avg segment duration	5.4s
Duration range	3.0s - 8.0s
Sample rate	24kHz mono
Format	Parquet with embedded audio

Emotion Distribution

Emotion	Count	Percentage
neutral	1894	83.8%
happy	188	8.3%
sad	148	6.5%
disgusted	22	1.0%
fearful	5	0.2%
angry	2	0.1%
surprised	2	0.1%

Speaker Distribution

Speaker	Segments
lex_fridman	551
pavel_durov	397
jeff_kaplan	359
norman_ohler	256
paul_rosolie	122
julia_shaw	122
jensen_huang	118
dan_houser	97
lars_brownworth	87
michael_levin	68
peter_steinberger	35
irving_finkel	31
david_kirtley	18

Dataset Structure

Column	Type	Description
audio	Audio	Waveform + sampling rate (24kHz)
text_verbatim	string	Verbatim transcript with fillers (umm, uh, [laughter], etc.)
text_verbatim_normalized	string	Verbatim text with numbers/abbreviations expanded (keeps fillers)
duration	float	Segment duration in seconds
snr	float	Signal-to-noise ratio (dB)
speaker_id	string	Speaker cluster ID (WavLM embeddings)
emotion	string	Speech emotion label (emotion2vec+ large, 9 categories)
pitch_mean	float	Mean F0 frequency (Hz)
pitch_std	float	F0 standard deviation (Hz)
energy_mean	float	Mean RMS energy
energy_std	float	RMS energy standard deviation
speaking_rate	float	Words per second
video_id	string	YouTube video ID
source_url	string	Source URL
start_time	float	Segment start time in source (seconds)
end_time	float	Segment end time in source (seconds)

Usage

from datasets import load_dataset

ds = load_dataset("somu9/emocean", split="train")

# Access a sample
sample = ds[0]
print(sample["audio"])       # {'path': ..., 'array': [...], 'sampling_rate': 24000}
print(sample["text"])
print(sample["emotion"])
print(sample["speaker_id"])

# Filter by emotion
happy = ds.filter(lambda x: x["emotion"] == "happy")

# Filter by speaker
spk0 = ds.filter(lambda x: x["speaker_id"] == "spk_0000")

Collection Pipeline

1. Download YouTube audio via yt-dlp
2. VAD segmentation (Silero VAD)
3. Quality filter — SNR > 25dB, clipping < 0.1%, music score < 0.5, boundary clip detection
4. Transcribe (Whisper large-v3)
5. Enrich — speaker embeddings (WavLM), prosody extraction, emotion classification (emotion2vec+ large)

License

CC-BY-4.0

---

Last updated: 2026-04-23 14:07 UTC