|
|
Detection of key information in emergency calls
Sarvaš, Marek ; Plchot, Oldřich (oponent) ; Schwarz, Petr (vedoucí práce)
Emergency calls are usually made under extremely stressful conditions, where callers often provide crucial information rapidly, making it difficult for emergency line agents to capture all details accurately. This can result in repeated questions about information that was already provided and cause delays in response times from emergency services. This work aims to mitigate this problem and potentially speed up the response of emergency services by deploying a neural network models for information extraction, specifically targeting the Named Entity Recognition (NER) task. This work explores various Transformer-based approaches for NER task, such as pre-trained encoder-only, encoder-decoder (sequence-2-sequence) and Large Language Models. The best models achieved state-of-the-art results on publicly available Czech NER datasets. In addition, new NER datasets were created from available recordings of real emergency calls and the corresponding metadata. The models were trained and evaluated on the created datasets successfully achieving reasonable performance in name and location extraction.
|
|
|
Interpretability of Neural Networks in Speech Processing
Sarvaš, Marek ; Mošner, Ladislav (oponent) ; Žmolíková, Kateřina (vedoucí práce)
With the growing popularity of deep neural networks, the lack of transparency caused by their black box representation is raising demand for their interpretability. The goal of this thesis is to gain new insights into deep neural networks in speech processing tasks. Specifically, gender classification task on AudioMNIST dataset and speaker classification task on filterbanks from VoxCeleb dataset using convolutional and residual neural network. Layer-wise relevance propagation was used for the interpretation of these neural networks. This method produced heatmaps highlighting features that contributed positively and negatively to the correct classification. As results of interpretation show, classifications were mainly based on lower frequencies in time. In the case of gender classification, I managed to find the model's high dependency on a small number of features. Using obtained information, I created an augmented training set that increased the model's robustness.
|
|
|
Interpretability of Neural Networks in Speech Processing
Sarvaš, Marek ; Mošner, Ladislav (oponent) ; Žmolíková, Kateřina (vedoucí práce)
With the growing popularity of deep neural networks, the lack of transparency caused by their black box representation is raising demand for their interpretability. The goal of this thesis is to gain new insights into deep neural networks in speech processing tasks. Specifically, gender classification task on AudioMNIST dataset and speaker classification task on filterbanks from VoxCeleb dataset using convolutional and residual neural network. Layer-wise relevance propagation was used for the interpretation of these neural networks. This method produced heatmaps highlighting features that contributed positively and negatively to the correct classification. As results of interpretation show, classifications were mainly based on lower frequencies in time. In the case of gender classification, I managed to find the model's high dependency on a small number of features. Using obtained information, I created an augmented training set that increased the model's robustness.
|
host ::
RSS.