Abstract
This paper describes a speaker diarization model based on target speaker voice activity detection (TS-VAD) using transformers. To overcome the original TS-VAD model's drawback of being unable to handle an arbitrary number of speakers, we investigate model architectures that use input tensors with variable-length time and speaker dimensions. Transformer layers are applied to the speaker axis to make the model output insensitive to the order of the speaker profiles provided to the TS-VAD model. Time-wise sequential layers are interspersed between these speaker-wise transformer layers to allow the temporal and cross-speaker correlations of the input speech signal to be captured. We also extend a diarization model based on end-to-end neural diarization with encoder-decoder based attractors (EEND-EDA) by replacing its dot-product-based speaker detection layer with the transformer-based TS-VAD. Experimental results on VoxConverse show that using the transformers for the cross-speaker modeling reduces the diarization error rate (DER) of TS-VAD by 11.3%, achieving a new state-of-the-art (SOTA) DER of 4.57%. Also, our extended EEND-EDA reduces DER by 6.9% on the CALLHOME dataset relative to the original EEND-EDA, achieving a new SOTA DER of 11.18%.
Introduction
Speaker diarization—the task of determining "who spoke when"—is notoriously difficult when multiple people talk at once. Two popular approaches exist, each with its own strengths and weaknesses. Target Speaker Voice Activity Detection (TS-VAD) is highly accurate but traditionally limited to a fixed number of speakers and sensitive to the order in which speaker profiles are provided. Conversely, End-to-End Neural Diarization with Encoder-Decoder Attractors (EEND-EDA) handles an arbitrary number of speakers but relies on a "dot-product" matching layer. This layer struggles with overlapping speech because it is mathematically difficult for a single audio frame's vector to point toward multiple speaker "attractors" simultaneously.
Transformer-based TS-VAD
The researchers introduced a major architectural shift by treating the data as a 3D tensor (Time × Speaker × Feature). Instead of concatenating speaker features into a flat vector, they apply Transformer layers along the speaker axis without positional encoding. This makes the model "order-invariant," meaning the output remains consistent regardless of the order in which speaker profiles are listed. These layers are alternated with BLSTM or Transformer layers along the time axis to capture temporal context.
Integration with EEND-EDA: EDA-TS-VAD
Recognizing that EEND-EDA’s dot-product layer was a performance bottleneck, the authors replaced it with the "Joint Speaker Detection" (JSD) module from their new TS-VAD. This allows the model to learn complex, non-linear relationships between audio frames and speaker attractors, significantly improving its ability to handle overlaps. The attractor-generation logic of EEND is used to feed the high-performance detection logic of TS-VAD.
Experimental Results
The proposed models were evaluated on two major benchmarks:
- VoxConverse: The Transformer-based TS-VAD achieved a Diarization Error Rate (DER) of 4.57%, an 11.3% relative improvement over the previous baseline. The improvement was most dramatic in sessions with many speakers (11+), where the error rate was reduced by nearly 17%.
- CALLHOME: The integrated EDA-TS-VAD model reached a DER of 11.18%, outperforming the original EEND-EDA (12.01%) and setting a new benchmark for the widely used training setting.
Conclusion
By removing positional encodings from speaker-axis Transformers, the model treats the set of speakers as a "bag of profiles," enhancing robustness. The hybridization of EEND and TS-VAD demonstrates that "end-to-end" and "profile-based" models are not mutually exclusive. Future work may involve refining temporal modeling within this 3D tensor framework, potentially incorporating positional encodings for the time-axis Transformer layers.
Summary updated