Emo-TTS Evaluation Datasets

This repository contains the datasets used for evaluating Emo-TTS, a training-free inference framework for high-arousal emotional speech synthesis.

📦 Datasets

Dataset	Description	Emotions	Speakers	Language	Format
HIED (ours)	High-Intensity Emotional Dataset for evaluation	Angry, Happy, Sad, Surprise	400 samples	EN	Parquet (HIED/)
ESD	Emotional Speech Dataset	Neutral, Happy, Sad, Angry, Surprise	20	EN / ZH	ZIP (ESD/)
EmoV-DB	Emotional Voices Database	Neutral, Amused, Angry, Sleepy, Disgusted	4	EN / FR	tar.gz (EmoV-DB/)
Expresso	High-quality expressive speech	8 read + 26 improvised styles	4	EN	Parquet (Expresso/)

📁 Repository Structure

.
├── HIED/                  # Our HIED benchmark (Parquet, loadable via 🤗 Datasets)
├── ESD/                   # Emotional Speech Dataset (ZIP archive)
├── EmoV-DB/               # EmoV-DB (per-speaker per-emotion tar.gz)
└── Expresso/              # Full Expresso dataset (Parquet)

🚀 Quick Start

Load HIED benchmark

from datasets import load_dataset

dataset = load_dataset("erminga/emo-tts", "HIED", split="test")
print(dataset[0])
# {'id': 'HIED_0000', 'audio': {...}, 'emotion': 'Angry', 'source_dataset': 'ESD', ...}

Download ESD

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="erminga/emo-tts",
    filename="ESD/Emotional_Speech_Dataset_ESD.zip",
    repo_type="dataset",
)

Download EmoV-DB

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="erminga/emo-tts",
    repo_type="dataset",
    allow_patterns="EmoV-DB/*",
    local_dir="./emo-tts-data",
)

Download Expresso

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="erminga/emo-tts",
    repo_type="dataset",
    allow_patterns="Expresso/*",
    local_dir="./emo-tts-data",
)

📊 Dataset Details

HIED (High-Intensity Emotional Dataset)

Our evaluation benchmark specifically designed to test TTS systems under high-arousal emotional conditions.

• Samples: 400 (100 per emotion)
• Emotions: Angry, Happy, Sad, Surprise
• Sources: ESD (354 samples), EmoV-DB (46 samples)
• Avg Duration: 3.85s
• Total Duration: ~0.43 hours

Field	Type	Description
id	string	Unique sample ID (e.g., HIED_0000)
audio	audio	Speech waveform
emotion	string	Emotion class (Angry/Happy/Sad/Surprise)
source_dataset	string	Source dataset (ESD/EmoV-DB)
speaker	string	Speaker identifier
rms_energy	float	RMS energy
f0_mean	float	Mean F0 (Hz)
f0_std	float	F0 standard deviation
f0_range	float	F0 range (Hz)
speaking_rate	float	Speaking rate
duration	float	Duration (seconds)

ESD (Emotional Speech Dataset)

• Paper: Zhou et al., 2022
• Speakers: 10 English + 10 Chinese
• Emotions: Neutral, Happy, Sad, Angry, Surprise
• Format: WAV files organized by speaker and emotion

EmoV-DB (Emotional Voices Database)

• Paper: Adigwe et al., 2018
• Source: OpenSLR-115
• Speakers: bea, jenie, josh, sam
• Emotions: Neutral, Amused, Angry, Sleepy, Disgusted

Expresso

• Paper: Nguyen et al., 2023
• Source: ylacombe/expresso
• Quality: 48kHz/24bit professional studio recordings
• Styles: 8 read speech styles + 26 improvised dialogue styles
• Duration: ~46 hours total

📜 License

• HIED: MIT
• ESD: CC BY-NC-SA 4.0
• EmoV-DB: CC BY 4.0
• Expresso: CC BY-NC 4.0

📖 Citation

If you use these datasets in your research, please cite our paper:

@inproceedings{emo-tts-2026,
  title={Emo-TTS: Rectifying Emotional Trajectories in Flow-Matching Speech Synthesis},
  author={},
  booktitle={ACL},
  year={2026}
}