SER_wav2vec2-large-xlsr-53_240304_fine-tuned_2

This model is a fine-tuned version of hughlan1214/SER_wav2vec2-large-xlsr-53_240304_fine-tuned1.1 on a Speech Emotion Recognition (en) dataset.

This dataset includes the 4 most popular datasets in English: Crema, Ravdess, Savee, and Tess, containing a total of over 12,000 .wav audio files. Each of these four datasets includes 6 to 8 different emotional labels.

This achieves the following results on the evaluation set:

• Loss: 1.0601
• Accuracy: 0.6731
• Precision: 0.6761
• Recall: 0.6794
• F1: 0.6738

Model description

The model was obtained through feature extraction using facebook/wav2vec2-large-xlsr-53 and underwent several rounds of fine-tuning. It predicts the 7 types of emotions contained in speech, aiming to lay the foundation for subsequent use of human micro-expressions on the visual level and context semantics under LLMS to infer user emotions in real-time.

Although the model was trained on purely English datasets, post-release testing showed that it also performs well in predicting emotions in Chinese and French, demonstrating the powerful cross-linguistic capability of the facebook/wav2vec2-large-xlsr-53 pre-trained model.

emotions = ['angry', 'disgust', 'fear', 'happy', 'neutral', 'sad', 'surprise']

Intended uses & limitations

More information needed

Training and evaluation data

70/30 of entire dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10

Training results

Framework versions

• Transformers 4.38.1
• Pytorch 2.2.1
• Datasets 2.17.1
• Tokenizers 0.15.2