Answers are grounded in this paper and cite their sources.
ty as SRS input features. Additional speaker-specific perceptual cues can be derived from behaviors and learned characteristics, such as vocabulary selection, accent, intonation style, and emotional aspects. Humans also use the speaker's sound signature similarity to known speakers to improve sound recognition precision. We need a new feature vector representation that compares a mark speaker's speech to a set of reference speaker's (codebook/dictionary). The speaker's utterance is encoded as cosine distance feature vectors (CDF). Back-end classifiers use SVMs (CDF-SVM). As a result, an SVM classifier with an intersection kernel captures the most acoustic similarities between target and reference speakers. Determining speaker discrimination is more important with reference speakers that are acoustically similar. Using CDF sparingly improves discriminative power by keeping only a few large values that correspond to the most similar reference speakers and setting all other elements to 0. On the core shorting condition of NIST's 2008 SRE databases, CDF-SVM outperforms SR systems using I-Vectors.
No takes yet. Share an insight, caveat, or question.
Reddy et al. (2024) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: