Search papers, labs, and topics across Lattice.
Stellenbosch University
3
0
3
Graph-based clustering can recover Zipfian distributions in unsupervised term discovery, outperforming traditional K-means methods.
Current lexicon evaluation methods are biased towards large clusters, but new metrics reveal a more accurate picture of lexicon quality.
Skip the multi-stage training: ZeroSyl offers a surprisingly simple, training-free way to extract syllable tokens directly from a frozen WavLM model, outperforming existing methods in spoken language modeling.
