---
datasets:
- narad/ravdess
language:
- en
metrics:
- f1
- accuracy
- recall
- precision
pipeline_tag: audio-classification
---

# Emotion Recognition in English Using RAVDESS and Wav2Vec 2.0

<!-- Provide a quick summary of what the model is/does. -->

This model extracts emotions from audio recordings. It was trained on RAVDESS, a dataset containing English audio recordings. The model recognises six emotions: anger, disgust, fear, happiness, sadness and surprise.

The model recreates the work of this [Greek emotion extractor](https://huggingface.co/m3hrdadfi/wav2vec2-xlsr-greek-speech-emotion-recognition/blob/main/README.md) using a pre-trained [Wav2Vec2](https://huggingface.co/jonatasgrosman/wav2vec2-large-xlsr-53-english) model to process the data. 


## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->


- **Adapted from:** [Emotion Recognition in Greek](https://huggingface.co/m3hrdadfi/wav2vec2-xlsr-greek-speech-emotion-recognition/blob/main/README.md)
- **Model type:** NN with CTC
- **Language(s) (NLP):** English
- **Finetuned from model:** [wav2vec2](https://huggingface.co/jonatasgrosman/wav2vec2-large-xlsr-53-english)


## How to Get Started with the Model

Use the code below to get started with the model.

[More Information Needed]

## Training Details

### Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

The RAVDESS dataset was split into training, validation and test sets with 60, 20 and 20 splits, respectively. 

### Training Procedure

The fine-tuning process was centred on four hyper-parameters: 
- the number of batches (4, 8),
- gradient accumulation steps (GAS) (2, 4, 6, 8),
- number of epochs (10, 20) and
- the learning rate (1e-3, 1e-4, 1e-5).

Each experiment was repeated 10 times.

## Evaluation

The set of hyper-parameters resulting in the best performance is: 4 batches, 4 GAS, 10 epochs and 1e-4 learning rate

## Testing

The model was retrained on the combined train and validation sets using the best hyper-parameter set. The performance on the test set has an average Accuracy and F1 scores of 84.84% (SD 2 and 2.08, respectively)


## Results

We retained the model providing the highest performance over the 10 runs.

| Emotion   | Accuracy | Precision |  Recall  |    F1    |
|-----------|:-------:|-----------:|---------:|---------:|
| Anger     |         |    96.55   |  87.50   |          |
| Disgust   |         |    90.91   |  93.75   |          |
| Fear      |         |    96.30   |  81.25   |          |
| Happiness |         |    93.10   |  84.38   |          |
| Sad       |         |    81.58   |  96.88   |          |
| Surprise  |         |    77.78   |  87.50   |          |
| Total     |  88.54  |    89.37   |  88.54   |  88.62   |


<!-- ## Citation [optional] -->

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

<!-- **BibTeX:**

[More Information Needed]

**APA:**

[More Information Needed] -->