Voice Cloning: Speaker Embedding Angle Map

Modern voice-cloning tools like ElevenLabs and Bark don't retrain a model for every new voice — they run a few seconds of reference audio through a frozen speaker encoder trained with a GE2E (Generalized End-to-End) loss, average the resulting embeddings, and hand that single vector to a synthesizer as a conditioning signal. Because cosine similarity between embeddings depends only on the angle between them, this simulator renders the entire embedding space as points on a circle: each point is one reference utterance at its embedding's angle, colour-coded by speaker. A training-progress slider replays how GE2E loss pulls each speaker's utterances onto a tight arc while pushing different speakers' centroid angles apart. Pick a target speaker and a number of few-shot reference clips, hit Clone, and watch a new golden angle get computed as their circular average — with live cosine-similarity bars showing exactly how a real speaker-verification system would judge whether the clone convincingly matches its target versus every other voice in the room.