TTS-AGI/emotion-attribute-conditioning-dacvae

Echo TTS - Emotion & Attribute Conditioning Dataset (DAC-VAE Latents)

Pre-bucketed speech dataset with DAC-VAE latent representations organized by 40 emotion categories and 13 vocal/audio attributes. Built for conditioning fine-tuning of Echo TTS and similar DiT-based TTS models.

Overview

• Total emotion samples: 163,271 (across 40 emotions, 10K cap per emotion)
• Total attribute samples: ~785K (across 13 attributes x 7 buckets, 10K cap per bucket)
• Format: WebDataset .tar files (312 total)
• Latent format: DAC-VAE 128-dim at 25 fps (variable length)
• Source: Filtered from 71.2M samples across 3 datasets

Source Datasets

1. TTS-AGI/podcast-tokenized-bg3.5-enj5 (481 tars)
2. TTS-AGI/podcast-tokenized-bg2.5-enj4.5 (6,496 tars)
3. TTS-AGI/emolia-hq-tokenized (10,491 tars)

Annotation Source

All emotion and attribute scores from LAION Empathic-Insight-Voice-Plus via emotion-annotations.

Data Format

Each sample in the tar files contains 3 files:

File	Shape	Description
{key}.npy	(frames, 128) float32	DAC-VAE latent (target speech, 25 fps)
{key}.ref.npy	(ref_frames, 128) float32	DAC-VAE latent (speaker reference)
{key}.json	-	Metadata with text, scores, speaker info

JSON Metadata Fields

Field	Type	Description
text	string	Transcript text
annotation_scores	dict (59 keys)	All emotion + attribute scores
caption	string	Natural language description of the audio
target_duration	float	Duration in seconds
context_duration	float	Speaker reference duration
latent_frames	int	Number of latent frames in .npy
ref_latent_frames	int	Number of frames in ref.npy
latent_dim	int	Always 128
speaker	string	Speaker ID
language	string	Language code
episode_id	int	Source episode identifier

Loading Example

import webdataset as wds
import numpy as np
import json

dataset = wds.WebDataset("emotions/Amusement/000000.tar")

for sample in dataset:
    key = sample["__key__"]
    latent = np.frombuffer(sample["npy"], dtype=np.float32).reshape(-1, 128)
    ref_latent = np.frombuffer(sample["ref.npy"], dtype=np.float32).reshape(-1, 128)
    metadata = json.loads(sample["json"])
    text = metadata["text"]
    scores = metadata["annotation_scores"]
    print(f"{key}: {latent.shape}, text='{text[:60]}...', amusement={scores.get('Amusement', 0):.2f}")

Directory Structure

emotions/
  {emotion_name}/         # 40 emotion directories
    000000.tar            # WebDataset tar (up to 4096 samples each)
    000001.tar
    ...

attributes/
  {attribute_name}/       # 13 attribute directories
    bucket_{i}_{lo}_to_{hi}/   # 7 range buckets per attribute
      000000.tar
      ...

Emotion Buckets (40 emotions)

Samples are routed to the emotion with the highest score, if that score >= 2.5. Each sample appears in at most one emotion bucket.

Emotion	Samples		Emotion	Samples
Interest	10,000		Pain	1,220
Hope_Enthusiasm_Optimism	10,000		Shame	1,215
Thankfulness_Gratitude	10,000		Infatuation	934
Amusement	10,000		Malevolence_Malice	874
Concentration	10,000		Doubt	507
Impatience_and_Irritability	10,000		Disgust	452
Contemplation	10,000		Awe	421
Anger	10,000		Contentment	376
Astonishment_Surprise	10,000		Contempt	348
Affection	10,000		Teasing	282
Confusion	9,756		Jealousy_&_Envy	203
Intoxication_Altered_States	9,375		Helplessness	194
Distress	6,182		Embarrassment	96
Fatigue_Exhaustion	5,672		Bitterness	74
Longing	4,452		Pleasure_Ecstasy	45
Relief	3,822		Sourness	27
Sadness	3,634
Elation	3,042
Triumph	2,001
Sexual_Lust	1,967
Emotional_Numbness	1,721
Fear	1,676
Pride	1,374
Disappointment	1,329

Total: 163,271 samples across 40 emotions

Attribute Buckets (13 attributes x 7 levels)

Each attribute is divided into 7 linear buckets. Samples can appear in multiple attribute buckets (one per attribute).

Age (0.0 - 6.0)

Bucket	Range	Samples
0	0.0 - 0.9	10,000
1	0.9 - 1.7	10,000
2	1.7 - 2.6	10,000
3	2.6 - 3.4	10,000
4	3.4 - 4.3	10,000
5	4.3 - 5.1	1,068
6	5.1 - 6.0	3

Arousal (0.0 - 4.0)

Bucket	Range	Samples
0	0.0 - 0.6	10,000
1	0.6 - 1.1	10,000
2	1.1 - 1.7	10,000
3	1.7 - 2.3	10,000
4	2.3 - 2.9	10,000
5	2.9 - 3.4	10,000
6	3.4 - 4.0	5,402

Confident vs. Hesitant (0.0 - 4.0)

Bucket	Range	Samples
0	0.0 - 0.6	10,000
1	0.6 - 1.1	10,000
2	1.1 - 1.7	10,000
3	1.7 - 2.3	10,000
4	2.3 - 2.9	10,000
5	2.9 - 3.4	10,000
6	3.4 - 4.0	1,376

Gender (-2.0 - 2.0)

Bucket	Range	Samples
0	-2.0 - -1.4	10,000
1	-1.4 - -0.9	10,000
2	-0.9 - -0.3	10,000
3	-0.3 - 0.3	10,000
4	0.3 - 0.9	10,000
5	0.9 - 1.4	10,000
6	1.4 - 2.0	10,000

Monotone vs. Expressive (0.0 - 4.0)

Bucket	Range	Samples
0	0.0 - 0.6	4,928
1	0.6 - 1.1	10,000
2	1.1 - 1.7	10,000
3	1.7 - 2.3	10,000
4	2.3 - 2.9	10,000
5	2.9 - 3.4	10,000
6	3.4 - 4.0	10,000

Serious vs. Humorous (0.0 - 4.0)

All 7 buckets at 10,000 samples.

Soft vs. Harsh (-2.0 - 2.0)

Bucket	Range	Samples
0	-2.0 - -1.4	349
1-5	...	10,000 each
6	1.4 - 2.0	1,176

Valence (-3.0 - 3.0)

All 7 buckets at 10,000 samples.

Warm vs. Cold (-2.0 - 2.0)

Bucket	Range	Samples
0	-2.0 - -1.4	420
1-6	...	10,000 each

Duration (1.0 - 30.0 seconds)

All 7 buckets at 10,000 samples.

Talking Speed (5.0 - 25.0 CPS)

All 7 buckets at 10,000 samples.

High-Pitched vs. Low-Pitched (0.0 - 4.0)

Bucket	Range	Samples
0	0.0 - 0.6	6
1-5	...	10,000 each
6	3.4 - 4.0	10

Submissive vs. Dominant (-3.0 - 3.0)

Bucket	Range	Samples
2	-1.3 - -0.4	162
3	-0.4 - 0.4	10,000
4	0.4 - 1.3	10,000
5	1.3 - 2.1	10,000
6	2.1 - 3.0	86

59 Annotation Score Dimensions

The annotation_scores field in each JSON contains all 59 dimensions from Empathic-Insight-Voice-Plus:

Emotions (40): Affection, Amusement, Anger, Astonishment_Surprise, Awe, Bitterness, Concentration, Confusion, Contemplation, Contempt, Contentment, Disappointment, Disgust, Distress, Doubt, Elation, Embarrassment, Emotional_Numbness, Fatigue_Exhaustion, Fear, Helplessness, Hope_Enthusiasm_Optimism, Impatience_and_Irritability, Infatuation, Interest, Intoxication_Altered_States_of_Consciousness, Jealousy_&_Envy, Longing, Malevolence_Malice, Pain, Pleasure_Ecstasy, Pride, Relief, Sadness, Sexual_Lust, Shame, Sourness, Teasing, Thankfulness_Gratitude, Triumph

Vocal Attributes (15): Age, Arousal, Authenticity, Background_Noise, Confident_vs._Hesitant, Gender, High-Pitched_vs._Low-Pitched, Monotone_vs._Expressive, Recording_Quality, Serious_vs._Humorous, Soft_vs._Harsh, Submissive_vs._Dominant, Valence, Vulnerable_vs._Emotionally_Detached, Warm_vs._Cold

Audio Quality (4): score_background_quality, score_content_enjoyment, score_overall_quality, score_speech_quality

Bucketing Strategy

Emotion routing: Each sample is assigned to the emotion with the highest score, if that score >= 2.5. Samples below threshold or with no strong emotion are excluded. Each sample appears in at most 1 emotion bucket.

Attribute routing: 7 linearly-spaced buckets per attribute. Each sample is placed into all matching attribute buckets (one per attribute dimension). Bucket size capped at 10,000 samples.

Pipeline Details

• Samples scanned: 71,242,881
• Processing time: ~15 hours
• Datasets processed: 17,468 tar files across 3 HuggingFace datasets

Citation

@misc{echo-conditioning-data,
  title={Echo TTS Emotion & Attribute Conditioning Dataset},
  url={https://huggingface.co/datasets/TTS-AGI/emotion-attribute-conditioning-dacvae},
  note={Bucketed DAC-VAE latent speech data with 59-dimensional emotion/attribute annotations},
}

Scores from LAION Empathic-Insight-Voice-Plus and emotion-annotations.