This repository contains the code used to perform the training and the analysis of the models described in the paper "Multimodal Emotion Recognition with High-level Speech and Text Features" accepted in the ASRU 2021 conference.
This repository is organized as follows:
-
aligner: contains a script used to extract phone alignment information using Gentle. This information is further used by the scripts in the
speechfolder. -
analysis: contains the scripts used to obtain the fused modalities (speech and text) results, and to get detailed results for each modality.
-
speech: contains the scripts to process the speech files, and to train and evaluate the speech emotion recognition model.
-
text: contains the scripts to extract text features using pre-trained Transformer-based models, and to train a text emotion recognition model over these features.
-
wav2vec: contains the scripts to extract wav2vec features from the speech files (these features are further used by the scripts in the
speechfolder).
We evaluated our method with the IEMOCAP dataset. Therefore, the scripts were written assuming it is possible to organize the data in 5 sessions, and according to the speaker's gender. To request access to the IEMOCAP dataset, and, to know more about this dataset, refer to the IEMOCAP webpage .
For information on how to run the scripts step-by-step, please refer to how_to.md, and, for information on the requirements to run the scripts of this repository, please refer to requirements.md.
The speech model found in the folder speech was inspired by the AutoVC model. For more information on AutoVC, check their github repository and paper.
For more information on the technical components of our method, on the results, and on the discussion of these results, please refer to our paper.