Graph Attention + Label Propagation (OCDGALP)

Updated 27 June 2026

OCDGALP is a framework that leverages graph attention to dynamically weigh node features and relationships within a graph.
It employs label propagation to iteratively refine cluster labels, ensuring consistency across local neighborhoods and global structures.
The approach is significant for applications requiring robust clustering in complex networks, offering improved scalability and accuracy.

Blockwise Neural Embedding + Constrained Clustering (EEND-Vector Clustering) is an integrated diarization paradigm that combines local neural diarization with global constrained clustering to solve speaker tracking and overlap handling in long-form, multi-speaker recordings. The EEND-VC framework leverages blockwise neural embedding extraction from short-segment EEND networks and enforces global speaker consistency using constrained clustering algorithms that respect intra-block prohibitions on speaker assignment. This methodology achieves state-of-the-art diarization error rates (DER) across multiple evaluation domains, balancing robustness and computational efficiency (Kinoshita et al., 2021, Serafini et al., 2023, Tawara et al., 2023, Pálka et al., 22 Oct 2025).

1. Conceptual Foundations and Motivation

The principal limitation of classical clustering-based diarization systems such as x-vector + PLDA + AHC/VBx lies in their inability to handle overlapping speakers and adapt to arbitrary recording lengths without explicit segment-level overlap labeling. Conversely, pure end-to-end neural diarization (EEND) models solve overlap and label-permutation issues in short segments but suffer from label permutation variability and memory scaling for long recordings. EEND-Vector Clustering hybridizes these approaches:

Local blocks (8–80 seconds) are processed independently by an EEND network, yielding both speaker-activity tracks and blockwise speaker embeddings.
All local embeddings are linked globally by constrained clustering, imposing “cannot-link” constraints that prohibit multiple embeddings from the same block being pooled into the same global speaker cluster.

This design maintains local overlap handling and global label consistency, efficiently scaling to real meeting data and open-domain conversational speech (Kinoshita et al., 2020, Kinoshita et al., 2021, Tawara et al., 2023).

2. Blockwise Neural Embedding: Network Architecture and Output

Block Segmentation and Neural Diarization

Long audio is segmented into non-overlapping (or lightly overlapped) blocks of duration $T$ (typically 15–80 s). Features per block (e.g., WavLM or log-Mel, dimension 23–1024) are fed into a Transformer-based EEND network, which yields:

Frame-level local speaker activity estimates $\widehat{y}_{t,i,s}\in[0,1]$
$S_\mathrm{Local}$ block-level d-dimensional speaker embeddings $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ (typically $d=256$ )

Each embedding is computed by weighted pooling of hidden states: $\widehat{\mathbf{e}}_{i,s} = \frac{\sum_{t=1}^{T} \widehat{y}_{t,i,s}\ \mathbf{h}_{t,i}}{\left\|\sum_{t=1}^{T} \widehat{y}_{t,i,s}\ \mathbf{h}_{t,i}\right\|}.$ Key architectural elements include:

Encoder: 2–6 Transformer layers, 4–8 heads, with hidden size $D=256-1024$
Local output streams: typically $S_\mathrm{Local}=3$
For overlap-heavy settings: integration of external embedding extractors (e.g., ECAPA-TDNN, ResNet-based x-vectors) under estimated masks further regularizes embedding consistency (Tawara et al., 2023).

Training employs a dual-objective permutation-invariant loss: $\mathcal{L} = (1-\lambda)\,\mathcal{L}_{\mathrm{diarization}} + \lambda\,\mathcal{L}_{\mathrm{embedding}}$

$\mathcal{L}_{\mathrm{diarization}}$ is PIT-based binary cross-entropy
$\widehat{y}_{t,i,s}\in[0,1]$ 0 is a contrastive or dictionary-based discriminative loss

This enforces that slot-wise embeddings are clusterable across blocks, and that neural speaker posterior streams remain permutable within blocks (Kinoshita et al., 2021, Kinoshita et al., 2020, Delcroix et al., 2023).

3. Constrained Clustering Algorithms

Global speaker consistency is established by clustering all blockwise embeddings $\widehat{y}_{t,i,s}\in[0,1]$ 1, subject to cannot-link constraints: within each block $\widehat{y}_{t,i,s}\in[0,1]$ 2, embeddings $\widehat{y}_{t,i,s}\in[0,1]$ 3, $\widehat{y}_{t,i,s}\in[0,1]$ 4 must not be assigned to the same global speaker.

Clustering Formulations:

Constrained Agglomerative Hierarchical Clustering (cAHC): Enforce cannot-link by setting inter-block-pair distances in $\widehat{y}_{t,i,s}\in[0,1]$ 5 to a large penalty $\widehat{y}_{t,i,s}\in[0,1]$ 6, and merge clusters only if no constraint is violated. Stopping can be at a fixed number of clusters $\widehat{y}_{t,i,s}\in[0,1]$ 7 or by threshold $\widehat{y}_{t,i,s}\in[0,1]$ 8.
Constrained Spectral Clustering (cSC): Cosine similarity matrix $\widehat{y}_{t,i,s}\in[0,1]$ 9 is masked on cannot-link pairs, Laplacian spectral clustering is applied, and $S_\mathrm{Local}$ 0-means is run in the eigenvector projection space (with eigengap or fixed $S_\mathrm{Local}$ 1) (Kinoshita et al., 2021).
COP-k-means: Classic $S_\mathrm{Local}$ 2-means objective

$S_\mathrm{Local}$ 3

subject to cannot-link, solved heuristically via constraint-respecting assignment and centroid alternation (Kinoshita et al., 2020, Tawara et al., 2023).

VBx-style Bayesian Clustering: Gaussian or spherical PLDA-based probabilistic cluster assignment, with or without HMM temporal smoothing. The multi-stream extension (MS-VBx) expands the standard VBx formalism to ensure streamwise cannot-link at each chunk (Delcroix et al., 2023, Pálka et al., 22 Oct 2025, Pálka et al., 23 Jun 2026).
SphereVBx: Variational Bayes in vMF mixtures (T-PSDA) for hyperspherical embeddings; parameter-free PF variants are strictly cosine-similarity clustering (Pálka et al., 23 Jun 2026).

Blockwise Integration: After clustering, local label-to-global assignment is resolved for every embedding, and diarization tracks are “stitched” according to the mapping, yielding globally consistent outputs (Kinoshita et al., 2021, Serafini et al., 2023, Tawara et al., 2023).

4. Inference Workflow and Algorithmic Pseudocode

The following pseudocode (modulo specific clustering backend) comprises the canonical EEND-VC pipeline (Kinoshita et al., 2021, Kinoshita et al., 2020):

$d=256$ 4

Key hyperparameters:

Blocklength $S_\mathrm{Local}$ 4 s
Local speaker slots $S_\mathrm{Local}$ 5
Embedding dimension $S_\mathrm{Local}$ 6
Activity threshold $S_\mathrm{Local}$ 7
For cAHC: penalty $S_\mathrm{Local}$ 8, linkage method “average” or “Ward” (Kinoshita et al., 2021, Delcroix et al., 2023, Tawara et al., 2023).

5. Empirical Evaluation and Performance Analysis

EEND-VC achieves substantial improvements over both vanilla EEND and classical VBx/AHC methods for diarization on long, overlapped, and open-domain recordings. Core findings include:

On CALLHOME (real English telephone data, 2–6 speakers): DER $S_\mathrm{Local}$ 9 (oracle speakers, cAHC), outperforming x-vector/PLDA systems by $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 0 absolute and previous EEND/EDA-EEND models by $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 1 (Kinoshita et al., 2021).
Chime-7 multi-channel meetings: EEND-VC with ECAPA embeddings and COP-kmeans clustering yields macro DER $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 2, the best among pure VC systems on the evaluation set (Tawara et al., 2023).
Compound “in-the-wild” benchmarks: EEND-VC with VBx-based clustering and short-segment filtering achieves DER $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 3 macro-average without dataset-level tuning (Pálka et al., 22 Oct 2025).
Experimental reviews confirm optimal DER and computational efficiency: RTF $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 4, RAM $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 5 GB, while requiring more labeled data than self-attentive EEND (Serafini et al., 2023).

Ablations show that constrained clustering (especially cAHC/COP-kmeans) outperforms unconstrained methods, particularly for $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 6 speakers. Blocklength $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 7 strongly mediates the tradeoff between embedding quality and permutation errors; best performance is observed for $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 8– $\widehat{\mathbf{e}}_{i,s}\in\mathbb{R}^d$ 9 s (Kinoshita et al., 2021).

6. Extensions: Temporal Modeling and Bayesian Clustering Backends

The VBx family of probabilistic clustering methods, originally Gaussian, has been generalized along multiple axes to improve global speaker assignments and robustness:

MS-VBx: Multistream extension of VBx for EEND-VC, modeling each chunk as generating multiple embeddings per HMM state, with transition probability $d=256$ 0 for speaker continuity. Inference is via VB EM using expected log-likelihoods for multistream Gaussian generative models, automatically enforcing cannot-link (Delcroix et al., 2023).
SphereVBx: Bayesian clustering with vMF mixtures and toroidal spherical discriminant analysis (T-PSDA), matching the intrinsic hyperspherical geometry of neural embeddings. PF (parameter-free) variants remove backend training, reducing heuristics and improving pipeline deployability (Pálka et al., 23 Jun 2026).
These models often integrate Hungarian- or permutation-based one-to-one assignment per chunk to handle cannot-link at the backend-inference level, as in MS-SphereVBx (Pálka et al., 23 Jun 2026).

7. Practical Considerations and Limitations

EEND-VC provides a robust, modular pipeline for diarization, but several factors impact deployment and performance:

Data Requirements: EEND-VC models require extensive diarization-annotated mixtures for effective training; performance degrades faster than classical systems when data is scarce (Serafini et al., 2023).
Chunk Configuration: Choice of blocklength and local speaker slots affects tradeoff between memory use, confusion errors, and clusterability (Kinoshita et al., 2021, Delcroix et al., 2023).
Clustering Complexity: COP-kmeans and cAHC scale polynomially but remain practical for moderate session lengths ( $d=256$ 1, $d=256$ 2); spectral methods require eigendecomposition of $d=256$ 3 affinity matrices.
Overlap Handling: The neural front-end provides robust overlap detection within local chunks. However, misalignment between masking and ECAPA embedding extraction can impact final cluster purity in overlap-heavy domains (Tawara et al., 2023).
Adaptivity: Late fusion and domain-specific self-supervised adaptation further improve robustness, especially in multi-microphone or domain-mismatched conditions (Tawara et al., 2023).

Future directions include tighter end-to-end joint optimization of EEND and the clustering backend, parameter-free Bayesian clusterers, and extension to spatial and visual embedding domains (Pálka et al., 22 Oct 2025, Pálka et al., 23 Jun 2026).

References:

(Kinoshita et al., 2020, Kinoshita et al., 2021, Serafini et al., 2023, Delcroix et al., 2023, Tawara et al., 2023, Pálka et al., 22 Oct 2025, Pálka et al., 23 Jun 2026)