Training-Free Affinity Fusion of Neural and Embedding-Based Speaker Diarization
Training-free affinity fusion combines neural and embedding diarization, reducing error on AMI and CALLHOME.
Training-Free Affinity Fusion combines a neural diarizer's speaker partition with an embedding-based acoustic affinity before one global clustering step. It does not require additional training, a shared embedding space, label alignment, or adoption of the neural system's speaker count. On AMI and CALLHOME it reduces diarization error versus both constituent systems, and on AMI it also improves speaker-attributed transcription. Ablations attribute most of the gain to the neural partition.
- TFAF builds a continuous affinity matrix from a neural speaker partition.
- Fusion needs no extra training, shared embeddings, or speaker-label alignment.
- AMI and CALLHOME show consistent diarization error-rate improvements.
- Neural partitions provide most of the gain; continuous affinities add more.
From the original article
1 min read at arxiv.orgSpeaker diarization systems based on speaker embeddings and neural diarization exploit complementary forms of speaker information, but their intermediate representations are not directly compatible. We introduce Training-Free Affinity Fusion (TFAF), which integrates the speaker structure inferred by a neural diarizer into an embedding-based diarization system.
By Yehoshua Dissen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.