Voice Activity Detection

[juanmc2005]

This PR adds a voice activity detection pipeline that is fully compatible with all of diart's features:

• Streaming, writing predictions to disk and plotting in real-time (only one label named "speech" will be displayed)
• Running Benchmark on a dataset and computing the detection error rate (instead of the diarization error rate)
• Running Benchmark in parallel
• Tuning hyper-parameters (in this case tau_active)
• Spinning up a websocket server

This is also implemented in the CLI interface and can be changed using the --pipeline argument, which requires the name of the pipeline to run. For the time being, the only possible options are SpeakerDiarization and VoiceActivityDetection. More to come in the future!

A significant refactoring was needed to squeeze in this feature, so backward compatibility with v0.7 is not guaranteed.
I also changed the names of some major classes so that they are more clear:

• BasePipeline becomes Pipeline
• BasePipelineConfig becomes PipelineConfig
• OnlineSpeakerDiarization becomes SpeakerDiarization (all pipelines are online)
• PipelineConfig becomes SpeakerDiarizationConfig (there are 2 pipelines now)
• RealTimeInference becomes StreamingInference ("real-time" depends on hardware, I'm more comfortable with "streaming")
• RealTimePlot becomes StreamingPlot (same as above)

Changelog

• Add VoiceActivityDetection pipeline with its VoiceActivityDetectionConfig
• Move base Pipeline, PipelineConfig and HyperParameter to diart.blocks.base
• Add --pipeline argument to CLI so the user can select a different pipeline to run, optimize, evaluate, etc.
• A Pipeline must be able to suggest an evaluation metric if none is provided
• Add metric: pyannote.metrics.BaseMetric parameter to Optimizer and Benchmark.__call__()
• Add direction: Literal["minimize", "maximize"] parameter to Optimizer
• Update GitHub link in setup.cfg

Notes on performance

Using pyannote/segmentation, duration=5s, step=0.5s, latency=5s and tau_active=0.507, the performance on AMI MixHeadset is:

Subset	Detection Error Rate	False Alarm	Missed Detection
test	5.9	3.2	2.7