Authors: Dr. N. V Rajasekhar Reddy, Sriyash Kulkarni, Thangella Sainikhil, Shreyas Vala
DOI Link: https://doi.org/10.22214/ijraset.2024.63859
Certificate: View Certificate
Speech is a powerful way to express our thoughts and feelings. It can give us valuable insights into human emotions. Speech emotion recognition (SER) is a crucial tool used in various fields like human-computer interaction (HCI), medical diagnosis, and lie detection. However, understanding emotions from speech is challenging. This research aims to address this challenge. It uses multiple datasets, including CREMA-D, RAVDESS, TESS, and SAVEE, to identify different emotional states. The researchers reviewed existing literature to inform their methodology. They used spectrograms and mel spectrograms extracted from speech data to capture important acoustic features for emotion recognition. The researchers used Convolutional Neural Networks (CNNs), a cutting-edge machine learning approach, to try and decipher the delicate emotional clues included in speech data. Accurate speech emotion recognition has important ramifications. They may result in more effective forensic investigations, better medical diagnosis, and enhanced human- computer interface experiences. With the potential to improve several sectors, this research advances the subject of emotional computing, which aims to comprehend the complex link between speech and emotion. CSS Concepts Clean User Interface (UI): Design a simple, intuitive UI with CSS for easy interaction. Responsive Design: Ensure UI adapts smoothly to different screen sizes using CSS media queries. Engaging Feedback: Use CSS animations for user feedback, like loading indicators. Consistent Branding: Apply CSS theming for a unified visual identity across the application.
I. INTRODUCTION
People interact by talking to each other. When they communicate, they express their emotions and feelings. By understanding the emotions of different speakers, we can tell if they are satisfied or not. This helps companies improve their services to customers, leading to the growth of the company. This idea is the basis for our project, "Speech Emotion Recognition Using Convolutional Neural Networks." Speech Emotion Recognition (SER) is an emerging technology in the field of Artificial Intelligence (AI). In recent times, SER has found applications in areas like Human-Computer Interactions, call centers, and forensics.
A. Spectrograms and Mel Spectrograms
Spectrograms and Mel Spectrograms are commonly used methods in Speech Emotion Recognition. Spectrograms are visual representations of sound waves that show the intensity and frequency of different sound components over time. They are created by applying a Fourier transform to the audio signals. Time is represented by the x-axis, frequency range by the y-axis, and audio signal amplitude by the colour. Mel spectrograms are similar to spectrograms, but they use a different frequency scale called the Mel scale. Mel spectrograms are widely used to extract features that are more representative of how humans perceive sound.
Both spectrograms and Mel spectrograms are useful tools for analyzing and visualizing speech signals. They provide valuable insights into the frequency content and changes over time, which can be important for various speech processing applications. This happens because the relationship between frequency and pitch is not linear. Sound signals are divided into small segments called frames, and the frequencies in each frame are analyzed. A technique called "mel scaling" is used to group similar frequencies and display them on a logarithmic scale. This better represents how humans perceive sound, as our ears are more sensitive to changes at lower frequencies. Next, a visual representation of the sound signal is displayed, with frequency on the vertical axis and time on the horizontal. More energy in the signal at that particular time and frequency is indicated by brighter colours. Below are some examples of both Mel Spectrograms and Spectrograms.
II. THE APPROACH IN BASE PAPER
Spectrograms and Mel spectrograms were created from brief audio recordings. The most effective method for extracting features was determined by analysing both spectrograms and mel spectrograms. Convolutional Neural Networks (CNN) are best suited for Speech Emotion Recognition (SER), according to prior studies. But resolving the researchers' concerns took precedence over creating the greatest possible model. It is impossible to determine whether a trained model is applicable in the actual world by using a single dataset. Consequently, data from several databases was used, something that had never been done previously.
III. EXISTING PROBLEMS
Overall, DeepEmotion represents an advanced approach to emotional recognition from speech, but it also faces some technical challenges that may need to be addressed for optimal performance and deployment.
In your current project, you can address the issue of limited generalization by using regularization techniques like dropout, batch normalization, and data augmentation. These methods can help prevent overfitting and encourage the model to learn more robust and versatile features from the spectrogram data.
Instead of designing complex CNN architectures from scratch, you could consider using pre-trained models like VGG16 and VGG19. These models have been pre-trained on large image datasets like ImageNet, which means they've already learned generic features that can be fine-tuned for speech emotion recognition tasks. By using transfer learning, you can benefit from these learned representations and reduce the computational complexity of training.
A. SERCNN
SERCNN is an existing syste•m that focuses on using convolutional neural networks (CNN) for spe•ech emotion recognition (SER). It e•mploys different CNN architecture•s, such as 1D-CNN or 2D-CNN, to process audio spectrograms repre•senting speech signals. The•se CNN models are traine•d on labeled emotion datase•ts to learn distinctive feature•s and accurately classify different e•motional states.
One drawback of SERCNN is that it may struggle to capture• long-term temporal depe•ndencies in spee•ch signals if it relies solely on CNN archite•ctures. Emotions are often e•xpressed dynamically over time•, and CNNs, which operate on fixed-size• windows, may not effectively capture• such temporal nuances. Another limitation is that training CNN-base•d models like SERCNN typically require•s a large amount of labeled data, which may not always be• readily available, espe•cially for specific emotion categorie•s or diverse speake•r demographics. This lack of data efficiency can be• a challenge.
Consider integrating convolutional neural networks (CNNs) and recurrent neural networks (RNNs) in your models to solve the constrained temporal environment. While RNNs can represent temporal dependencies, which aids the model in capturing subtle emotional expressions across time, CNNs are capable of extracting spatial characteristics from spectrums.
To improve data e•fficiency, try using data augmentation techniques like time stretching, pitch shifting, and noise• injection on your spectrogram and mel spectrogram data. Additionally, you can leverage transfer learning by fine-tuning pretraine•d VGG16 and VGG19 models on your speech e•motion recognition (SER) datasets. By initializing the mode•ls with weights learned from ge•neric image data, you can train them e•ffectively with smaller amounts of labeled speech data, enhancing data efficiency and model performance.
IV. IMPLEMENTATION
A. Dataset Details
The CREMA-D datase•t is a multimodal dataset consisting of 7443 video clips featuring 91 actors. The• emotions portrayed in the datase•t include neutral, happiness, ange•r, disgust, fear, and sadness. This dataset was cre•ated using crowdsourcing, with 2443 raters providing the e•motion labels, making it a reliable re•source.. In contrast, the RAVDESS dataset has 1440 audio files that are solely for speech out of 7356 files that include song and speech data. RAVDESS portrays the following emotions: peaceful, joyful, sad, furious, fearful, shocked, and disgusted. For the purpose of teaching neural networks to recognise various emotional expressions, both datasets offer insightful data. It is mentioned in the text that not every emotion from the RAVDESS dataset was used. Since they are uncommon in other datasets and the RAVDESS dataset is too tiny to be used exclusively for deep neural network (DNN) training, the emotions of surprise and tranquilly were eliminated. Because of this, there are somewhat uneven amounts of examples for each emotion—particularly for the "neutral" feeling, which includes a mere 48 files.
The following six emotions are represented by the TESS (Toronto Emotional Speech Set) dataset: numbness, pleasant surprise, rage, disgust, fear, and happiness. There are 2800 audio files in this dataset, which includes recordings of two female speakers who are 26 and 64 years old. However, since only two actors were participating in the study, using the TESS dataset alone for CNN model training would be nearly impossible because half of the dataset would need to be reserved for testing. Because of this, other datasets including CREMA-D, RAVDESS, and SAVEE are often used in conjunction with the TESS dataset. The "Surrey Audio-Visual Expressed Emotion" dataset is abbreviated as SAVEE. It features audio, visual, and audiovisual recordings of four English male actors portraying seven distinct emotions: fear, happiness, sadness, disgust, rage, and so on.
With 90 instances of the "neutral" emotion—double the amount of the other emotions—the dataset exhibits an imbalanced distribution. It is not large enough on its own for artificial neural network (ANN) training, but it can be paired with the larger RAVDESS dataset, which has less examples of "neutral" emotions.
Convolutional Neural Networks (CNNs) are a type of deep learning model, which are an extension of artificial neural networks. They have proven to be highly effective in areas like image recognition and audio/video analysis.
The process starts by taking an image as input and passing it through the core components of a CNN. The key ide•a behind CNNs is to apply filters, called kernels, to the input image in order to extract the most relevant features for the specific task at hand.
The main layers in a CNN include:
By carefully structuring these layers, CNNs can effectively learn to recognize complex patterns in visual data, making them a powerful tool for various computer vision applications.
B. Convolution
A crucial layer in deep learning models is convolution. It is employed to identify significant aspects, such as images, in the input data. It uses filters, sometimes referred to as kernels, to achieve this. These filters create a feature map that indicates the locations of specific features by performing a dot product with the underlying pixel values.
The activation function plays a vital role in convolution. It introduces non-linearity, meaning the input is not directly proportional to the output. This is important, as it allows the model to learn more complex patterns beyond basic classification. Padding is sometimes used in convolution to avoid losing information at the edges, as the feature map size can get reduced after the convolution operation.
Pooling is another important layer that follows convolution. It downsamples the feature maps, reducing the computational load and memory requirements. Pooling extracts the most salient features, helping the model focus on the most important information.
Convolution and pooling, taken together, allow deep learning models to efficiently identify and extract significant features from incoming data.Different forms of pooling exist. Max pooling, sum pooling, and average pooling are a few of them.
C. Flattening
The Flattening layer is usually added right be•fore the fully connected layer. This layer is necessary to convert the 2D matrix into a single-column matrix, which can be• fed as input to the fully connected layer.
D. Alexnet
With eight layers, Alexnet is a potent deep learning network. Five convolutional layers are used at first, then max-pooling layers, and finally three completely connected layers. A sizable image database called ImageNet is used to train this network.
The Alexnet layer layers are:
Because of its complex and deep network architecture, Alexnet is exceptionally good at computer vision tasks like picture classification.
V. ARCHITECTURE
We first downloaded the CREMA-D, RAVDESS, SAVEE, and TESS datasets from Kaggle. Then, we organized the datasets based on the emotions they represented. Next, we used the Librosa library to extract spectrograms and mel spectrograms from the audio files. We split these spectrograms and mel spectrograms into testing and training sets.
The CNN model demonstrated an accuracy of 74.69% in recognizing 4 emotions and 62.17% in recognizing 6 emotions when using spectrograms. We also split the mel spectrograms into training and testing sets, which resulted in accuracies of 75.87% and 64.06% respectively for 4 and 6 emotions. This suggests that mel spectrograms performed better than regular spectrograms in identifying emotions. Additionally, we experimented with different architectures like AlexNet, VGG-16, and VGG-19, and found that VGG-19 provided the highest accuracy in recognizing both 4 and 6 emotions. Our research emphasises how crucial it is to partition datasets appropriately for AI models\' testing and training. While many studies report impressive speech emotion recognition (SER) results, interdependent data splitting is a problem that is frequently disregarded. Because of this, it may be difficult to check and compare the results directly, particularly if the software is not publicly available to the research community. We carried out experimental comparisons to highlight the need of dataset splitting techniques in order to allay this worry. In summary, this study shows that mel-spectrograms are an effective feature extraction method for convolutional neural networks (CNNs) in speech emotion recognition tasks. The benefit of mel-spectrograms is evident in our quantitative visualisation, even though spe•ctrograms are still used in the literature. Going ahead, it is essential
[1] Zielonka, M.; Piastowski, A.; Czy?ewski, A.; Nadachowski, P.; Operlejn, M.; Kaczor, K. \"Recognition of Emotions in Speech Using Convolutional Neural Networks on Different Datasets.\" Electronics 2022, 11, 3831. [2] Abeer Ali Alnuaim, Mohammed Zakariah, Prashant Kumar Shukla, Aseel Alhadlaq, Wesam Atef Hatamleh, Hussam Tarazi, R. Sureshbabu, Rajnish Ratna, \"Human-Computer Interaction for Recognizing Speech Emotions Using Multilayer Perceptron Classifier\", Journal of Healthcare Engineering, vol. 2022, Article ID 6005446, 12 pages, 2022. [3] Singh, A., Srivastava, K. K., & Murugan, H. \"Speech Emotion Recognition Using Convolutional Neural Network (CNN).\" International Journal of Psychosocial Rehabilitation, Vol. 24, Issue 08, 2020. [4] Anvarjon, T.; Mustaqeem; Kwon, S. \"Deep-Net: A Lightweight CNN-Based Speech Emotion Recognition System Using Deep Frequency Features.\" Sensors 2020, 20, 5212. [5] F. Andayani, L. B. Theng, M. T. Tsun and C. Chua, \"Hybrid LSTM-Transformer Model for Emotion Recognition From Speech Audio Files,\" in IEEE Access, vol. 10, pp. 36018-36027, 2022. [6] L. Yunxiang and Z. Kexin, \"Design of Efficient Speech Emotion Recognition Based on Multi Task Learning,\" in IEEE Access, vol. 11, pp. 5528-5537, 2023. [7] M. B. Er, \"A Novel Approach for Classification of Speech Emotions Based on Deep and Acoustic Features,\" in IEEE Access, vol. 8, pp. 221640-221653, 2020. [8] K. V. Krishna, N. Sainath and A. M. Posonia, \"Speech Emotion Recognition using Machine Learning,\" 2022 6th International Conference on Computing Methodologies and Communication (ICCMC), Erode, India, 2022, pp. 1014-1018. [9] Eyben, F., Wöllmer, M., & Schuller, B. \"Opensmile: the Munich versatile and fast open-source audio feature extractor.\" Proceedings of the 18th ACM international conference on Multimedia. ACM, 2010. [10] Schuller, B., Steidl, S., Batliner, A., Vinciarelli, A., Scherer, K., Ringeval, F., ... & Burkhardt, F. \"The INTERSPEECH 2013 computational paralinguistics challenge: social signals, conflict, emotion, autism.\" Proceedings INTERSPEECH, 2013. [11] Goodfellow, I., Bengio, Y., & Courville, A. \"Deep learning.\" MIT press, 2016. [12] LeCun, Y., Bengio, Y., & Hinton, G. \"Deep learning.\" Nature, 521(7553), 436-444, 2015. [13] Deng, J., Dong, W., Socher, R., Li, L. J., Li, K., & Fei-Fei, L. \"Imagenet: A large-scale hierarchical image database.\" IEEE conference on computer vision and pattern recognition, 2009. [14] He, K., Zhang, X., Ren, S., & Sun, J. \"Deep residual learning for image recognition.\" Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. [15] Simonyan, K., & Zisserman, A. \"Very deep convolutional networks for large-scale image recognition.\" arXiv preprint arXiv:1409.1556, 2014. [16] Hinton, G., Deng, L., Yu, D., Dahl, G. E., Mohamed, A. R., Jaitly, N., ... & Kingsbury, B. \"Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups.\" IEEE Signal Processing Magazine, 29(6), 82- 97, 2012. [17] Graves, A., & Schmidhuber, J. \"Framewise phoneme classification with bidirectional LSTM and other neural network architectures.\" Neural Networks, 18(5-6), 602-610, 2005. [18] Kingma, D. P., & Ba, J. \"Adam: A method for stochastic optimization.\" arXiv preprint arXiv:1412.6980, 2014. [19] Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., ... & Zheng, X. \"TensorFlow: Large-scale machine learning on heterogeneous systems.\" Software available from tensorflow.org, 2015. [20] Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., ... & Desmaison, A. \"PyTorch: An imperative style, high- performance deep learning library.\" Advances in Neural Information Processing Systems, 32, 8024-8035, 2019.
Copyright © 2024 Dr. N. V Rajasekhar Reddy, Sriyash Kulkarni, Thangella Sainikhil, Shreyas Vala. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Paper Id : IJRASET63859
Publish Date : 2024-08-01
ISSN : 2321-9653
Publisher Name : IJRASET
DOI Link : Click Here