They use pyannote/speaker-diarization. I tried that, but it wasn't accurate enough for my purposes. Made a confusion matrix with voice samples from The Simpsons

They use pyannote/speaker-diarization. I tried that, but it wasn't accurate enough for my purposes. Made a confusion matrix with voice samples from The Simpsons characters. It looked... well, confused. Am using a mix now of speaker embeddings and other signals (end of sentence, pause before/after, etc.). As you can see in the demos, it already works well for interview situations. It's when there are 3+ speakers and they talk over each other that the system gets confused.