Cascade method for voice-based speaker search using cross-model comparisons

We address voice-based speaker search in a large gallery of voices. Modern models pretrained with self-supervised learning, such as wav2vec 2.0, provide high search accuracy but require substantial computation for extracting speaker embeddings. To reduce this cost, we propose a cascaded method: the gallery is enrolled using a fast student model, dWhisper-ASV, while queries and candidate verification are performed by a teacher model, wav2vec 2.0 XLSR-1B. The alignment of embedding spaces enables cross-model cosine similarity. On ClovaCall, Speech-MASSIVE, and a proprietary Russian corpus, RusIVR_test, the cascade nearly matches teacher-only F1-score: 77.21 % vs. 77.24 %, 83.05 % vs. 83.16 %, and 92.26 % vs. 92.63 %. Only 1.5 %, 0.2 %, and 0.11 % of records reach the second stage and the enrollment speeds up eightfold.