mexca: Capture emotion expressions from multiple modalities in videos

Adds average speaker embeddings and improved speaker diarization. Also increases the performance of data processing. Provides an advanced example notebook for extending the standard MEXCA pipeline.

Added

• The SpeakerAnnoation class has a new attribute speaker_average_embeddings containing the average embeddings for each detected speaker
• The SpeakerIdentifier has a new argument to explicitly set the device its run on (by default CPU)
• The SpeakerIdentifier.apply() method has a new show_progress argument to enable progress bars for detected speech segments and embeddings
• A new notebook on customizing and extending the MEXCA pipeline (examples/example_custom_pipeline_components.ipynb)
• Two new recipes for applying the standard MEXCA pipeline and postprocessing the extracted features (recipes/)
• The Pipeline.apply() method has a new merge argument to disable merging features from different modalities; this is useful when customizing a pipeline
• A new logo (thanks to Ji Qi)
• Documentation on how to use mexca with GPU and CUDA support
• notebook has been added as a dependency for the demo installation
• scikit-learn has been added as an explicit dependency (previously dependency of py-feat)

Changed

• pyannote.audio has been upgraded to version 3.0.0; this required adding the following dependencies:
  • torch >= 2.0.0
  • onnxruntime-gpu on Windows and Linux
  • onnxruntime on MacOS
  • torchaudio on MacOS
• torch has been upgraded to version 2.0.0 for all components requiring it
• The SpeakerIdentifier component uses the pyannote/speaker-diarization-3.0 model by default
• pandas has been replaced by polars; the Multimodal.features attribute now stores a polars.LazyFrame instead of a pandas.DataFrame; this speeds up postprocessing and merging for large data sets

Removed

• py-feat has been removed as a dependency