Emotion Classification from Audio

A HuggingFace pipeline for emotion classification from audio using SpeechBrain ECAPA embeddings and SVM.

Model Details

• Architecture: SpeechBrain ECAPA-TDNN embeddings (192-dim) + SVM classifier
• Training Data: RAVDESS dataset
  • 7 emotion classes: angry, disgust, fearful, happy, neutral/calm, sad, surprised
  • Professional actors (24 speakers)
  • 5-fold cross-validation
• Performance:
  • RAVDESS 5-fold cross-validation: 86.24% accuracy
• Audio Processing:
  • Input format: Any audio file format supported by soundfile
  • Automatically converted to: 16kHz, mono, single channel, 256 Kbps

Installation

You can install the package directly from GitHub:

pip install git+https://github.com/griko/voice-emotion-classification.git

Usage

from voice_emotion_classification import EmotionClassificationPipeline

# Load the pipeline
classifier = EmotionClassificationPipeline.from_pretrained(
    "griko/emotion_7_cls_svm_ecapa_ravdess"
)

# Single file prediction
result = classifier("path/to/audio.wav")
print(result)  # ["angry"] or ["disgust"] or ["fearful"] or ["happy"] or ["neutral/calm"] or ["sad"] or ["surprised"]

# Batch prediction
results = classifier(["audio1.wav", "audio2.wav"])
print(results)  # ["angry", "disgust"]

Limitations

• Model was trained on actor voices from RAVDESS dataset
• Performance may vary on:
  • Different audio qualities
  • Different recording conditions
  • Multiple simultaneous speakers
  • Non-actor voices
  • Different languages or accents

Citation

If you use this model in your research, please cite:

@misc{koushnir2025vanpyvoiceanalysisframework,
      title={VANPY: Voice Analysis Framework}, 
      author={Gregory Koushnir and Michael Fire and Galit Fuhrmann Alpert and Dima Kagan},
      year={2025},
      eprint={2502.17579},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2502.17579}, 
}

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.

Acknowledgments

• RAVDESS dataset for providing the training data
• SpeechBrain team for their excellent speech processing toolkit