This is the official github page for the paper:
Gullal S. Cheema, Judi Arafat, Chiao-I Tseng, John A. Bateman, Ralph Ewerth, and Eric Müller-Budack. 2024. Identification of Speaker Roles and Situation Types in News Videos. In Proceedings of the 2024 International Conference on Multimedia Retrieval (ICMR '24). Association for Computing Machinery, New York, NY, USA, 506–514. https://doi.org/10.1145/3652583.3658101
-
Create environment for training/evaluation
conda env create -f environment_main.yml
-
Install transnetv2, clip and download spacy model [For feature extraction]
-
Clone TransNetV2
cd TransNetV2 python setup.py install
-
pip install git+https://github.com/openai/CLIP.git
-
python -m spacy download de_core_news_lg
-
-
Additionally setup environment for diarization and feature extraction
conda env create -f environment_diarization.yml
- Install whisperx:
pip install git+https://github.com/m-bain/whisperX.git@4cb167a225c0ebaea127fd6049abfaa3af9f8bb4
- Install whisperx:
- Splits, features and speaker turn clip names used in the paper are available in
dataset/
- Drop an email for access to speaker turn clips and full videos
- Structure of feature pickle files (
bild_features_segmentbased.pkl
orbild_features_windowbased.pkl
):- Dictionary with keys as clip names (e.g., '20220105_Corona_Regeln_Unsere_Freiheit_gerät_0RM8KQi3Muk_1')
- Each speaker type clip name key contains a dictionary with the following keys:
- 'feature': Array containing the feature vector
- 'label_0': Speaker role label for level 0 (anchor, reporter, external)
- 'label_1': Speaker role label for level 1 (anchor, reporter, expert, politician, layperson, other)
- 'start': Start time of the speaker turn in seconds
- 'end': End time of the speaker turn in seconds
- Each situation type clip name key contains a dictionary with the following keys:
- 'feature': Array containing the feature vector
- 'label': News situation label (talking-head, voiceover, interview, commenting, speech)
- 'start': Start time of the speaker turn in seconds
- 'end': End time of the speaker turn in seconds
- Example:
{ '20220105_Corona_Regeln_Unsere_Freiheit_gerät_0RM8KQi3Muk_1': { 'feature': array([7.32914681e-02, 6.23861488e-03, ...], # Feature vector 'label_0': 2, 'label_1': 3, 'start': 1.579, 'end': 29.7 }, '20220120_Omikron_Welle_Diese_Impfpflicht_ist_pWZDF3rJ744_231': { ... }, ... }
- Each speaker type clip name key contains a dictionary with the following keys:
- Dictionary with keys as clip names (e.g., '20220105_Corona_Regeln_Unsere_Freiheit_gerät_0RM8KQi3Muk_1')