Hugging Face's logo

Datasets:
mcernak
/
EARS-EMO-OpenACE

Tasks:
Audio Classification
Audio-to-Audio
Languages:
English
ArXiv:
Tags:
emotional-speech
audio-quality
perceptual-evaluation
codec-evaluation
mushra
visqol
License:
Dataset card Files
xet
Community
1
EARS-EMO-OpenACE / README.md
mcernak's picture
mcernak
Improve dataset card: Add paper and code links, add 'benchmark' tag (#1)
31acd60 verified
preview code
|
raw
history blame contribute delete
5.94 kB
metadata
language:
  - en
license: mit
size_categories:
  - <1K
task_categories:
  - audio-classification
  - audio-to-audio
tags:
  - emotional-speech
  - audio-quality
  - perceptual-evaluation
  - codec-evaluation
  - mushra
  - visqol
  - polqa
  - benchmark

EARS-EMO-OpenACE: A Full-band Coded Emotional Speech Quality Dataset

Paper | Code

Dataset Description

This dataset contains full-band coded emotional speech samples at 16 kbps with human perceptual quality ratings and objective quality metrics. It is designed for research in audio quality assessment, emotion recognition, and codec evaluation.

Key Features

  • Coded audio sample using Opus, LC3, LC3Plus, and EVS codecs
  • Full-band emotional speech samples (48kHz, 24-bit), total 252 files:
  • Size of 6 speakers x 6 emotions x 4 codecs = 144 scored audio files
  • Size of 6 speakers x 6 emotions x 2 MUSHRA anchors = 72 scored files
  • Size of 6 speakers x 6 emotions x 1 reference = 36 reference files
  • Human MUSHRA ratings (perceptual quality) obtained from 36 listeners
  • Reference and coded audio pairs available for each emotion and codec
  • Computed VISQOL objective quality scores
  • Computed POLQA objective quality scores

Applications

  • Audio codec evaluation
  • Perceptual quality modeling
  • Emotion-aware audio processing
  • Quality metric validation

Dataset Structure 

EARS-EMO-OpenACE/
├── metadata.csv              # Main dataset metadata
├── dataset_summary.json      # Dataset statistics
├── README.md                 # This file
└── [speaker]/                # Speaker directories
    └── emo_[emotion]_freeform/   # Emotion directories
        ├── reference.wav         # Reference audio
        ├── [codec].wav           # Coded audio files
        └── ...

Metadata Schema

Column Description Scale/Range
dataset Dataset identifier EARS-EMO-OpenACE
speaker Speaker ID p102, p103, p104, p105, p106, p107
emotion Emotional expression Anger, Ecstasy, Fear, Neutral, Pain, Sadness
codec Audio codec used EVS, LC3, LC3Plus, Opus, LowAnchor, MidAnchor
reference_file Path to reference audio Relative path
distorted_file Path to coded audio Relative path
reference_mushra_rating Human quality rating for reference 0-100 (MUSHRA scale)
distorted_mushra_rating Human quality rating for coded audio 0-100 (MUSHRA scale)
mushra_rating_difference Quality degradation Reference - Distorted
visqol_score VISQOL objective quality score 1-5 (higher = better)
polqa_score POLQA objective quality score 1-5 (higher = better)

Codec Information

  • Reference: EARS source file
  • EVS: Enhanced Voice Services codec
  • LC3: Low Complexity Communication Codec
  • LC3Plus: Enhanced version of LC3
  • Opus: Open-source audio codec
  • LowAnchor: Low-quality anchor (lp3500) - Human ratings only
  • MidAnchor: Mid-quality anchor (lp7000) - Human ratings only

Quality Metrics

MUSHRA Ratings

  • Scale: 0-100 (higher = better quality)
  • Method: Multiple Stimuli with Hidden Reference and Anchor
  • Raters: Trained human listeners
  • Reference: Original uncompressed audio (typically ~95-100)

VISQOL Scores

  • Scale: 1-5 (higher = better quality)
  • Method: Virtual Speech Quality Objective Listener
  • Type: Objective perceptual quality metric
  • Coverage: Available for most codecs (excluding anchors)

POLQA Scores

  • Scale: 1-5 (higher = better quality)
  • Method: Perceptual Objective Listening Quality Assessment
  • Standard: ITU-T P.863
  • Coverage: Available for most codecs (excluding anchors)

Usage Examples

Load Dataset 

import pandas as pd

# Load metadata
metadata = pd.read_csv('metadata.csv')

# Filter by emotion
anger_samples = metadata[metadata['emotion'] == 'Anger']

# Filter by codec
opus_samples = metadata[metadata['codec'] == 'Opus']

# Get high-quality samples (MUSHRA > 80)
high_quality = metadata[metadata['distorted_mushra_rating'] > 80]

Audio Loading 

import librosa

# Load audio file
audio_path = metadata.iloc[0]['distorted_file']
audio, sr = librosa.load(audio_path, sr=None)

Correlation Analysis

The following table shows Pearson correlations between objective metrics and human MUSHRA ratings:

Metric Description Correlation (r) p-value Sample Size Significance
VISQOL Virtual Speech Quality Objective Listener 0.7034 8.3513e-23 144 ***
POLQA Perceptual Objective Listening Quality Assessment 0.7939 2.9297e-32 143 ***

Significance levels: *** p<0.001, ** p<0.01, * p<0.05, n.s. = not significant

Note: Correlations computed excluding anchor codecs (lp3500/lp7000) which only have human ratings.

Citation

If you use this dataset in your research, please cite the following paper:

@inproceedings{OpenACE-Coldenhoff2025,
  author={Coldenhoff, Jozef and Granqvist, Niclas and Cernak, Milos},
  booktitle={ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={{OpenACE: An Open Benchmark for Evaluating Audio Coding Performance}}, 
  year={2025},
  pages={1-5},
  keywords={Codecs;Speech coding;Audio coding;Working environment noise;Benchmark testing;Data augmentation;Data models;Vectors;Reverberation;Speech processing;audio coding;benchmarks;deep learning;speech processing},
  doi={10.1109/ICASSP49660.2025.10889159}
}

License

MIT License

Contact

Milos Cernak, milos.cernak at ieee dot org

August 1, 2025