---
language: multilingual
license: apache-2.0
datasets:
- voxceleb2
- timit
libraries:
- speechbrain
- librosa
tags:
- age-estimation
- speaker-characteristics
- speaker-recognition
- audio-regression
- voice-analysis
---

# Age Estimation Model

This model combines the SpeechBrain ECAPA-TDNN speaker embedding model with an ANN regressor to predict speaker age from audio input. The model uses ECAPA embeddings and Librosa acoustic features, trained on both VoxCeleb2 and TIMIT datasets.

## Model Performance Comparison

We provide multiple pre-trained models with different architectures and feature sets. Here's a comprehensive comparison of their performance:

| Model | Architecture | Features | Training Data | Test MAE | Best For |
|-------|-------------|----------|---------------|-----------|----------|
| VoxCeleb2 SVR (223) | SVR | ECAPA + Librosa (223-dim) | VoxCeleb2 | 7.88 years | Best performance on VoxCeleb2 |
| VoxCeleb2 SVR (192) | SVR | ECAPA only (192-dim) | VoxCeleb2 | 7.89 years | Lightweight deployment |
| TIMIT ANN (192) | ANN | ECAPA only (192-dim) | TIMIT | 4.95 years | Clean studio recordings |
| Combined ANN (223) | ANN | ECAPA + Librosa (223-dim) | VoxCeleb2 + TIMIT | 6.93 years | Best general performance |

You may find other models [here](https://huggingface.co/griko).

## Model Details
- Input: Audio file (will be converted to 16kHz, mono, single channel)
- Output: Predicted age in years (continuous value)
- Features:
  - SpeechBrain ECAPA-TDNN embedding [192 features]
  - Additional Librosa features [31 features]
- Regressor: Artificial Neural Network optimized through Optuna
- Performance:
  - Combined test set: 6.93 years Mean Absolute Error (MAE)

## Features
1. SpeechBrain ECAPA-TDNN embeddings (192 dimensions)
2. Librosa acoustic features (31 dimensions):
   - 13 MFCCs
   - 13 Delta MFCCs
   - Zero crossing rate
   - Spectral centroid
   - Spectral bandwidth
   - Spectral contrast
   - Spectral flatness

## Training Data
The model was trained on a combination of datasets:
- VoxCeleb2:
  - YouTube interview recordings
  - Age data from Wikidata and public sources
  - Voice activity detection applied
- TIMIT:
  - Studio-quality recordings
  - Original age annotations
- All audio preprocessed to 16kHz, mono
## Installation

```bash
pip install git+https://github.com/griko/voice-age-regression.git#egg=voice-age-regressor[full]
```

## Usage

```python
from age_regressor import AgeRegressionPipeline

# Load the pipeline
regressor = AgeRegressionPipeline.from_pretrained(
    "griko/age_reg_ann_ecapa_librosa_combined"
)

# Single file prediction
result = regressor("path/to/audio.wav")
print(f"Predicted age: {result[0]:.1f} years")

# Batch prediction
results = regressor(["audio1.wav", "audio2.wav"])
print(f"Predicted ages: {[f'{age:.1f}' for age in results]} years")
```

## Limitations
- Model was trained on a mix of YouTube interviews and studio recordings recordings
- Performance may vary on different audio qualities or recording conditions
- Age predictions are estimates and should not be used for medical or legal purposes
- Age estimations should be treated as approximate values, not exact measurements

## Citation
If you use this model in your research, please cite:
```bibtex
@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}, 
}
```