VAD emotion scoring for audio files

Model versioning

There are three main versions of the model architecture:

• V1: Uses the MelEncoder as described in this paper and the ERes2NetV2 Speaker Recognition Model.
• V1s: Uses same backbone as V1, but predicts mean value and standart deviation for VAD-scores.
• V2: Utilizes the BUD-E Whisper model to extract feature embeddings and applies a simple MLP head to predict a vector of VAD scores.
• V3: Also uses the BUD-E Whisper model for feature embeddings, but employs a more complex two-head system to predict the means and standard deviations of VAD scores. This approach better captures the variability in emotion annotations.
• V3s: Similar to V3, but uses GELU activation instead of PReLU in the MLP heads. It has almost the same performance as V3, but is more efficient in terms of memory usage.
• V4: Uses Emotion2Vector large model as a backbone, followed by learnable query and attention. Uses two heads architecture as vXs. The emo2vec realization for realization is taken from FunASR toolkit

For more detailed information about the architecture, check the architecture overview document.

Weights

Model weights are available on Hugging Face. At the moment (18.08.2025) the following weights are available:

1. V3 weight in root directory of the repository. The model was trained on the Laion's Got Talent (Enhanced Flash Annotations and Long Captions) dataset with Variant 1 annotation approach, which is described in detail in the annotation methodology document.

2. V3s weight in the scorer_v3s_16_final directory. This model was trained on the same dataset but with Variant 2 annotation approach, which is described in the same document. The model uses GELU activation instead of PReLU in the MLP heads, as well as projection layers applied to encoder output embeddings. This model is more memory efficient and has almost the same performance (loss wise) as V3. Model configuration is available in the scorer_v3s_16_final/configs/vad_train_config.yaml file.

Update Also V3s weights trained on the same data as V4 available in the same repository in v3s_balanced directory. This time all checkpoints are available for testing.

3. V4 weights in the v4_balanced directory. Overall, subjectively, this variant perform the best especially when it comes to determined Dominance scores.

Configuration files

For simplicity reason, configuratuon is done through a YAML file. Check the "vad_train_config.yaml" file for example. All parameters that must be adjusted are marked with comments. Parameters that are NOT marked with comments must stay as they are, since they are model specific and should not be changed.

Data formatting

The annotation file is expected to be .csv that uses ; as separator. The file should contain the following columns:

• full_path - the full path to the audio file
• pleasure_mean - the mean pleasure score for the audio file
• pleasure_std - the standard deviation of the pleasure score for the audio file
• arousal_mean - the mean arousal score for the audio file
• arousal_std - the standard deviation of the arousal score for the audio file
• dominance_mean - the mean dominance score for the audio file
• dominance_std - the standard deviation of the dominance score for the audio file
• verified_emotion - the verified emotion label for the audio file based on scores means. It used for balancing the emotion distribution in the dataset in case some emotions are overrepresented or underrepresented.

For more detailed information about the annotation process check the document here or exmaine this notebook here.

How to abtain VAD scores for your audio files

If your dataset contains categorical emotion labels, you can use the mapping table described on page 15 of the original paper to convert them to the continuous pleasure, arousal and dominance means and standard deviations.

Also, you can download extracted mapping table from Google Drive here.

Which model to choose?

Choose between V4 and V3s, considering the following characteristics:

1. V3s tends to misinterpret high arousal as high valence and shows a bias toward lower valence scores.
2. V4 tends to smooth arousal means toward zero, reducing the magnitude of extreme values.
3. V4 is better when it comes to determining extreme expressivness in dominance values.
4. V4 is more robust to semantic leakage, making it more stable when semantic content influences emotional predictions.

Roadmap

• [+] Upload V3 weights to Hugging Face
• [+] Upload V3s weights to Hugging Face
• [+] Upload V3s weights trained on a balanced dataset to Hugging Face
• [+] Upload V4 weights trained on a balanced dataset to Hugging Face
• [+] Upload formatted emotion to VAD score mapping table
• Add more documentation on how to use the model
• [-] Implement and train V1 model with mean and standard deviation heads (Model showed a very poor performance)

Credits

1. GPT SoVITS for initial ideas and code for certain modules.
2. LaionAI for the Laion's Got Talent (Enhanced Flash Annotations and Long Captions) dataset as well as the BUD-E Whisper model.
3. MER2025
4. FanASR