🎙️ 2D Few-Shot Speaker Enrolment: Embedding Averaging, Cosine Scoring and Equal Error Rate
Score histograms for few-shot speaker enrolment: N noisy embeddings averaged into a template, cosine scores for the true speaker and impostors, the equal error rate and a closed-form normal approximation, with a projection of the shots, verified by Monte Carlo (synthetic embeddings).
⚙ Under the hood
Score histograms for few-shot speaker enrolment: N noisy embeddings averaged into a template, cosine scores for the true speaker and impostors, the equal error rate and a closed-form normal approximation, with a projection of the shots, verified by Monte Carlo (synthetic embeddings).
Speaker VerificationSpeaker EmbeddingCosine SimilarityEqual Error RateFew-Shot LearningNoise AveragingHigh-Dimensional GeometryVoice Biometrics
2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install