Artificial Neural Network (ANN) models, specifically Convolutional Neural Networks (CNN), were applied to extract emotions based on spectrograms and mel-spectrograms. This study uses spectrograms and mel-spectrograms to investigate which feature extraction method better represents emotions and how big the differences in efficiency are in this context. The conducted studies demonstrated that mel-spectrograms are a better-suited data type for training CNN-based speech emotion recognition (SER). The research experiments employed five popular datasets: Crowd-sourced Emotional Multimodal Actors Dataset (CREMA-D), Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS), Surrey Audio-Visual Expressed Emotion (SAVEE), Toronto Emotional Speech Set (TESS), and The Interactive Emotional Dyadic Motion Capture (IEMOCAP). Six different classes of emotions were used: happiness, anger, sadness, fear, disgust, and neutral. However, some experiments were prepared to recognize just four emotions due to the characteristics of the IEMOCAP dataset. A comparison of classification efficiency on different datasets and an attempt to develop a universal model trained using all datasets were also performed. This approach brought an accuracy of 55.89% when recognizing four emotions. The most accurate model for six emotion recognition was trained and achieved 57.42% accuracy on a combination of four datasets (CREMA-D, RAVDESS, SAVEE, TESS). What is more, another study was developed that demonstrated that improper data division for training and test sets significantly influences the test accuracy of CNNs. Therefore, the problem of inappropriate data division between the training and test sets, which affected the results of studies known from the literature, was addressed extensively. The performed experiments employed the popular ResNet18 architecture to demonstrate the reliability of the research results and to show that these problems are not unique to the custom CNN architecture proposed in experiments. Subsequently, the label correctness of the CREMA-D dataset was studied through the employment of a prepared questionnaire.

In the medical field, text annotation involves categorizing clinical and biomedical texts with specific medical categories, enhancing the organization and interpretation of large volumes of unstructured data. This process is crucial for developing tools such as speech recognition systems, which help medical professionals reduce their paperwork. It addresses a significant cause of burnout reported by up to 60% of medical staff. However, annotating medical texts in languages other than English poses unique challenges and necessitates using advanced models. In our research, conducted in collaboration with Gdańsk University of Technology and the Medical University of Gdańsk, we explore strategies to tackle these challenges. We evaluated the performance of various tools and models in recognizing medical terms within a comprehensive vocabulary, comparing these tools’ outcomes with annotations made by medical experts. Our study specifically examined categories such as ‘Drugs’, ‘Diseases and Symptoms’, ‘Procedures’, and ‘Other Medical Terms’, contrasting human expert annotations with the performance of popular multilingual chatbots and natural language processing (NLP) tools on translated texts. The conclusion drawn from our statistical analysis reveals that no significant differences were detected between the groups we examined. This suggests that the tools and models we tested are, on average, similarly effective—or ineffective—at recognizing medical terms as categorized by our specific criteria. Our findings highlight the challenges in bridging the gap between human and machine accuracy in medical text annotation, especially in non-English contexts, and emphasize the need for further refinement of these technologies.

This research investigates the perceived naturalness of synthesized speech in the context of Polish medical terminology, a critical factor for applications such as voice-enabled medical dialogue systems. We conducted a comparative analysis of three speech synthesis models: SpeechGen, ElevenLabs, and a version of ToucanTTS fine-tuned on a specialized corpus of Polish medical recordings. The evaluation employed objective measures, the NISQA metric, and subjective assessments through Mean Opinion Score (MOS) surveys. Our findings indicate that SpeechGen and ElevenLabs produce synthesized speech that closely rivals the naturalness of human speech, as evidenced by both NISQA scores and MOS ratings. In contrast, despite improvements, the fine-tuned ToucanTTS model did not achieve comparable levels of perceived naturalness. Notably, participants occasionally rated the advanced synthesized speech as more natural than human speech recorded in non-studio environments, underscoring the potential of these technologies in real-world applications. This study emphasizes the significance of naturalness in enhancing user experience, particularly in specialized linguistic domains. It provides insights into speech synthesis's current capabilities and limitations for less-resourced languages like Polish.