• Convolutional part includes 3 ConvBlock for extracting local information from mel spectrogram input.
  ConvBlock : Conv1d -> ConvLN -> GELU -> Dropout
• The returned output of Convolutional part goes through Maxpool1d and LayerNormalization.
• GRU part extracts global information bidirectionally. GRU has an advantage in handling variable-length input.
• BahdanauAttention is applied at the end of GRU. This enables the model to know where to pay attention.

• train/test dataset with a proportion of 8:2.
• Stratified sampling from the emotion class distribution
• Mean test accuracy of 84.17(± 2.36)% in Stratified 5-fold cross-validation

• The start and end points are obtained from power of mel spectrogram.
• This process helps the model to learn features of the voiced region. There cannot be any emotion in the silent regions(before/after speech), hence they are excluded from analysis.

• Speech audio-only files from Full Dataset
• With 8 emotion classes: neutral, calm, happy, sad, angry, fearful, surprise, and disgust
  • View more dataset information here
• Without augmentation or extra-dataset

• Loss function: LabelSmoothingLoss (source)
• Optimizer: adabelief_pytorch.AdaBelief (source)
• Scheduler: torch.optim.lr_scheduler.ReduceLROnPlateau
• Please check train.py for more detailed hyper-parameters.

Best Test Accuracy: 87.15% / Loss: 0.86552

Train Accuracy: 0.98881(± 0.00367)
Train Loss: 0.63214(± 0.00590)
Test Accuracy: 0.85262(± 0.01182)
Test Loss: 0.86931(± 0.01257)

• ❗️The mean accuracy during the last 10 epochs may be more reliable to us than a simple best accuracy.

• We can see the confusion matrix(8 x 8). Rows represent the actual emotion classes while columns represent the predicted classes.
• Normalized confusion matrix shows recall for each emotion class.
• Accuracy metric is the same as WAR(Weighted Average Recall) since I did stratified sampling.

• ❗️ We can see that the the model is most confused about the happy class.

Due to the small size of the dataset, there was a concern that the hold-out output might be biased, so I attempted to perform 5-fold cross-validation.

[Fold 1] Best Test Accuracy: 0.84375 / Loss: 0.97918
[Fold 2] Best Test Accuracy: 0.80903 / Loss: 1.04934
[Fold 3] Best Test Accuracy: 0.85764 / Loss: 0.87293
[Fold 4] Best Test Accuracy: 0.87500 / Loss: 0.90747
[Fold 5] Best Test Accuracy: 0.82292 / Loss: 0.98223
[5-Fold CV Best Test Accuracy] max: 0.87500 min: 0.80903 mean: 0.84167 std: 0.02361

• ❗️ The mean test accuracy is approximately 84%, which is slightly (almost 3%p) lower than the hold-out result, but overall showing good performance.
• ❗️ One drawback is that there is significant variation in test performance across folds, which could be due to shuffling of the data, but it could also be a reproducibility issue with the optimizer.

• The worst-performing and best-performing output are compared through the following figure.
• The confusion matrices for all folds can be found in /output/cv5/img/

• ❗️ There was a significant gap between the highest and lowest values, especially the accuracy for "happy" class was notably low.

• With EmoCatcher, I achieved a test accuracy of 87.15% on a hold-out dataset and mean test accuracy of 84.17% under 5-fold CV.
• I have personally tested this on the CPU. According to the PyTorch documentation, it is not guaranteed that output will be reproducible across different devices. Since I haven't conducted many experiments yet, I think more experiments should be carried out to find consistent parameters to increase reproducibility of the above output.
• There was overfitting of the train data, which was probably due to small dataset size. (Insufficient data size may have made performance estimates noisy.) Given the circumstances, it was a good enough performance.
• When I listened and guessed it, my accuracy was about 60%.😅 I think it was difficult to detect emotional characteristics contained in the speech due to cultural differences. (I'm Korean.)
• So next time, I will challenge Korean Speech Emotion Recognition.

The following figures are the output of the best model on the hold-out: output/holdout/model/best_model_0.8715_0.8655.pth

• We can check which parts the attention mechanism focuses on for each class through the following figure.

• Plot title format: true emotion class/predicted emotion class(predicted probability x 100 %)

If you want to use this code, please cite as follows:

@misc{hwang9u-emocatcher,
 author = {Kim, Seonju},
 title = {emocatcher},
 year = {2023},
 publisher = {GitHub},
 journal = {GitHub repository},
 howpublished = {\url{https://github.com/hwang9u/emocatcher}},
}