emolia-balanced-5M-subset

A balanced ~5.26M-sample subset of laion/Emolia (80.5M speech samples), packaged as WebDataset-compatible tar shards for direct use in training pipelines.

---

How this subset was filtered

Samples were selected if they met either of two criteria:

1. Emotion thresholds

Each sample carries 40 emotion annotation scores (from the Emonet taxonomy) in its metadata. A sample qualifies for an emotion bucket if its score for that emotion meets or exceeds a per-emotion threshold. Thresholds were computed from the full Emolia distribution such that each emotion bucket targets ~100,000 samples.

Up to 100,000 samples per emotion × 40 emotions were collected (with overlap — one sample can qualify under multiple emotions).

2. Speaker diversity (centroid coverage)

Each sample's 128-dimensional WavLM timbre embedding was assigned to its nearest centroid from a set of 3,000 speaker centroids (k-means pruned, centroids_pruned.npy). Up to 1,000 samples per centroid were collected to ensure broad speaker diversity across rare voice types.

Deduplication

After extraction, sample IDs were checked globally — 259 cross-shard duplicates were removed, leaving 5,256,683 unique samples.

---

Emotion bucket fill

Emotion	Threshold	Samples
Affection	≥ 1.50	100,000
Amusement	≥ 2.00	100,000
Anger	≥ 1.50	100,000
Astonishment/Surprise	≥ 2.00	96,332 ¹
Awe	≥ 1.00	100,000
Bitterness	≥ 1.00	100,000
Concentration	≥ 2.50	100,000
Confusion	≥ 2.00	100,000
Contemplation	≥ 2.00	100,000
Contempt	≥ 1.50	100,000
Contentment	≥ 1.50	100,000
Disappointment	≥ 1.50	100,000
Disgust	≥ 1.00	100,000
Distress	≥ 1.50	100,000
Doubt	≥ 1.00	100,000
Elation	≥ 2.00	100,000
Embarrassment	≥ 1.00	100,000
Emotional Numbness	≥ 2.00	100,000
Fatigue/Exhaustion	≥ 1.50	100,000
Fear	≥ 1.50	100,000
Helplessness	≥ 1.00	100,000
Hope/Optimism	≥ 2.50	100,000
Impatience and Irritability	≥ 2.00	100,000
Infatuation	≥ 1.00	100,000
Interest	≥ 2.50	100,000
Intoxication/Altered States	≥ 1.50	100,000
Jealousy & Envy	≥ 1.00	100,000
Longing	≥ 1.50	100,000
Malevolence/Malice	≥ 1.50	100,000
Pain	≥ 1.00	100,000
Pleasure/Ecstasy	≥ 1.50	100,000
Pride	≥ 1.50	100,000
Relief	≥ 1.50	100,000
Sadness	≥ 1.50	100,000
Sexual Lust	≥ 1.00	100,000
Shame	≥ 1.00	100,000
Sourness	≥ 1.00	100,000
Teasing	≥ 1.50	100,000
Thankfulness/Gratitude	≥ 3.00	100,000
Triumph	≥ 1.50	100,000

¹ Astonishment/Surprise reached only 96,332 — the rarest emotion in the full 80.5M-sample dataset; all available qualifying samples were included.

---

Dataset format

Shards are plain tar files containing sequential .mp3 / .json pairs:

emolia-000000.tar
  000000.mp3
  000000.json
  000001.mp3
  000001.json
  ...
  004999.mp3
  004999.json
emolia-000001.tar
  ...

• 1,052 shards × 5,000 samples = 5,256,683 samples total
• Audio: MP3, original quality from Emolia
• Metadata: original Emolia JSON with wavelm_timbre_embedding stripped (saves ~50% JSON size), plus __emolia_id__ field

JSON fields

Each .json file is the original Emolia sample metadata plus:

Field	Meaning
__emolia_id__	Original Emolia sample ID
id	Same as __emolia_id__ (original field)
language	BCP-47 language code
duration	Audio duration in seconds
dnsmos	DNSMOS audio quality score
speaker	Speaker ID
emotion_annotation	Dict of 40+ *_best emotion scores
characters_per_second	Speech rate

---

Loading in PyTorch

import webdataset as wds

dataset = (
    wds.WebDataset("data/emolia-{000000..001051}.tar")
    .decode("torch")          # decodes .mp3 with torchaudio
    .to_tuple("mp3", "json")
)

for audio, meta in dataset:
    # audio: torch.Tensor   (decoded waveform)
    # meta:  dict           (Emolia metadata)
    emotion_scores = meta["emotion_annotation"]
    ...

Or load a single shard directly:

import tarfile, json

with tarfile.open("data/emolia-000000.tar") as tar:
    members = {m.name: m for m in tar.getmembers()}
    for i in range(5000):
        key = f"{i:06d}"
        audio = tar.extractfile(members[key + ".mp3"]).read()
        meta  = json.load(tar.extractfile(members[key + ".json"]))

---

Source dataset

• Source: laion/Emolia — 80.5M speech samples with 40-dimensional emotion annotations (Emonet taxonomy)
• Emotion annotations: continuous regression scores from a fine-tuned audio model, stored under emotion_annotation.*_best
• Speaker embeddings: 128-dim WavLM timbre embeddings (used for centroid assignment, stripped from this subset's JSON)
• Thresholds derived from: 86% of full dataset (~69M samples), with a ×1.16 scale factor to extrapolate to 100k/emotion target

---

License

Inherits the license of laion/Emolia. Please refer to the source dataset for usage terms.