Target Speaker Voice Activity Detection with Transformers and Its Integration with End-To-End Neural Diarization

Dongmei Wang, Xiong Xiao, Naoyuki Kanda, Takuya Yoshioka, Jian Wu

TL;DR

Researchers at Microsoft have overhauled Target Speaker Voice Activity Detection (TS-VAD) by replacing fixed-length layers with Transformers, allowing the model to handle an unlimited number of speakers regardless of their order. By integrating this flexible architecture into End-to-End Neural Diarization (EEND-EDA), they achieved new state-of-the-art results on the VoxConverse and CALLHOME benchmarks.

What Problem Does This Paper Address?

Speaker diarization—the task of determining "who spoke when"—is notoriously difficult when multiple people talk at once. Two popular approaches exist:

  1. TS-VAD: Highly accurate but traditionally limited to a fixed number of speakers and sensitive to the order in which speaker profiles are provided.
  2. EEND-EDA: An end-to-end approach that handles an arbitrary number of speakers but relies on a "dot-product" matching layer. This layer struggles with overlapping speech because it’s mathematically difficult for a single audio frame's vector to point toward multiple speaker "attractors" simultaneously.

The authors aim to break these limitations by making TS-VAD flexible and using its robust joint-detection capabilities to fix the overlap bottleneck in EEND-EDA.

Methods

The paper introduces two major architectural shifts:

  • Transformer-based TS-VAD: Instead of concatenating speaker features into a flat vector (which fixes the speaker count), the authors treat the data as a 3D tensor (Time × Speaker × Feature). They apply Transformer layers along the speaker axis without positional encoding. This makes the model "order-invariant"—it doesn't matter if Speaker A is listed first or fifth. They alternate these with BLSTM or Transformer layers along the time axis to capture temporal context.
  • EDA-TS-VAD: The authors recognized that EEND-EDA’s dot-product layer was a performance bottleneck. They 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 the audio frames and the speaker attractors, significantly improving its ability to handle overlaps.

Datasets

  • VoxConverse: Used to evaluate the TS-VAD improvements. It contains multi-speaker YouTube videos with 1–21 speakers per session.
  • CALLHOME: Used to evaluate the integrated EEND-EDA model. This is a classic dataset of recorded telephone conversations.
  • VoxCeleb1 & 2: Used for pre-training the speaker embedding extractors and simulating large-scale conversational data for initial model training.

Key Insights and Ideas

  • Permutation Invariance is Power: By removing positional encodings from the speaker-axis Transformers, the model treats the set of speakers as a "bag of profiles." This is more robust than previous "multi-target" models where the output could change if you simply swapped the order of the inputs.
  • The Dot-Product Bottleneck: The authors provide a geometric intuition for why EEND-EDA struggled: a dot product measures proximity in vector space. If two people are talking, the audio embedding has to be "close" to two different speaker vectors at once, which is a difficult constraint to satisfy. The TS-VAD JSD module bypasses this by using a neural network to "decide" activities based on the whole speaker set.
  • Hybridization Wins: The paper shows that "end-to-end" models (EEND) and "profile-based" models (TS-VAD) aren't mutually exclusive. You can use the attractor-generation logic of EEND to feed the high-performance detection logic of TS-VAD.

Key Results

  • VoxConverse SOTA: The Transformer-based TS-VAD achieved a Diarization Error Rate (DER) of 4.57%, an 11.3% relative improvement over the previous baseline.
  • Handling Crowds: The improvement was most dramatic in sessions with many speakers. For sessions with 11+ speakers, the new model reduced the error rate by nearly 17% compared to the original TS-VAD.
  • CALLHOME SOTA: 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.

Limitations and Future Work

While the Transformer-based TS-VAD is flexible, it still benefits from high-quality initial speaker profiles. In the TS-VAD experiments, these profiles were generated using a separate clustering step (AHC).

The authors also noted that while Transformers worked best for the speaker axis, BLSTMs still slightly outperformed Transformers on the time axis in some configurations. They suggest this might be because their time-axis Transformer didn't use positional encoding, which is usually vital for sequence modeling. Future work could involve refining the temporal modeling within this 3D tensor framework.

Summary updated

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 with a similar model size, achieving a new SOTA DER of 11.18% under a widely used training data setting.
BibTeX
@inproceedings{wang2023target,
  author = {Dongmei Wang and Xiong Xiao and Naoyuki Kanda and Takuya Yoshioka and Jian Wu},
  title = {{Target Speaker Voice Activity Detection with Transformers and Its Integration with End-To-End Neural Diarization}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=FaRZUA2gPU},
}

Loading related content...

Figure