Joint Factor Analysis

• Front-end Factor Analysis for Speaker Verification
• Joint Factor Analysis (JFA)
• eature extractor to learn a low-dimensional speaker representation for Speaker Verification, which is also used to model session and channel effects/variabilities
• In this new space, a given speech utterance is represented by a new vector named total factors (called the identity-vector or the “i-vector”)
• The i-vector is thus a feature that represents the characteristics of the frame-level Features’ distributive pattern
• [dimensionality reduction](dimensionality reduction.md) of the GMM supervector (although the GMM supervector is not extracted when computing the i-vector)
• extracted in a similar manner with the eigenvoice adaptation scheme or the JFA technique
• extracted per sentence
• Support-Vector-Machine-based system that uses the cosine kernel to estimate the similarity between the input data
• Cosine Similarity as the final decision score
• removed the SVM from the decision proces
• no speaker enrollment
• EER
• MinDCF
• NIST 2008 Speaker Recognition Evaluation dataset
• Up until d-vectors, the state-of-the-art Speaker Verification systems were based on the concept of i-vectors