When a fake voice clip surfaces, the useful forensic question is often not only "is this synthetic?" but "what made it?". An analyst may need to know whether two clips came from the same generator, or whether a clip came from a diffusion-based system with a GAN vocoder (the component that turns an intermediate spectrogram into a waveform). Generators are built from shared parts, so their family tree is real. Many text-to-speech systems reuse the same vocoder, and systems differ first by broad design and then by specific components.
We propose to learn forensic embeddings in hyperbolic space. This is a curved geometry in which available room grows exponentially with distance from a center point, so a tree can be laid out without its leaves crowding together. In this geometry the distance from the center can encode how specific the attribution is. A clip from a known generator should land far out, near its leaf. A clip from a never-seen generator should land at an intermediate radius, under the deepest ancestor it truly shares with known systems. The result is a single embedding that supports pairwise "same source?" verification and also graded attribution that stops at the right level of the tree instead of forcing a wrong leaf.
Core Claim
Euclidean forensic embeddings flatten the generator taxonomy. They therefore fail in a specific, measurable way on generators that are new but partially familiar. We claim that a low-dimensional hyperbolic embedding, trained with an explicit generator-component taxonomy, will do three things. It will match or beat the best Euclidean verification system at a fraction of the embedding dimension. It will place unseen generators at a radius that tracks how much of their ancestry is known. It will attribute such clips to the correct internal node of the tree more often than a Euclidean model given the same taxonomy labels. The third claim is the one that separates geometry from supervision, and the experiments are designed around it.
Relation to Prior Work
The motivating evidence comes from Audio Deepfake Verification. That work defines open-set verification (deciding whether two clips share a generative source, even an unseen one) and builds Audity, a dual-branch model. One branch is w2v-BERT 2.0 for prosody and phoneme-transition structure; the other is CAM++ for low-level vocoder artifacts. Audity reports that systems sharing blocks such as HiFi-GAN vocoders cluster together in its Euclidean space. This is exactly the latent tree we want to model explicitly. Audity scores pairs with Euclidean or cosine similarity and uses no taxonomy. We keep its two-branch front end and change the geometry and the supervision.
The closest hyperbolic work is Poin-HierNet, which learns multiple Poincaré-ball prototypes and organizes them hierarchically for binary real-versus-fake detection (1.40% EER on ASVspoof 2021 DF, 4.91% on In-the-Wild). Its hierarchy is latent and unlabeled, its output is a real/fake decision, and it never evaluates attribution or verification. We do not compete on binary detection. We use known component labels to shape the tree and evaluate whether the geometry yields correct partial attribution. HierCon argues for hierarchy across SSL layers in Euclidean space. Its hierarchy is over network depth rather than generator ancestry, so it is complementary rather than competing.
Three face anti-spoofing papers constrain the design: Hyperbolic Face Anti-Spoofing, Hyp-OC, and HPDR. All three gain from Poincaré-ball attack hierarchies. All three report boundary instability that needs feature or gradient clipping, and HPDR notes sensitivity to curvature. We therefore adopt the Lorentz (hyperboloid) model from Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry. That paper shows closed-form geodesics, no boundary singularity, and a 10-dimensional embedding that beats 200-dimensional Euclidean ones on WordNet.
The idea that unfamiliar inputs drift toward the center, and that this signals uncertainty, comes from Balanced Hyperbolic Embeddings Are Natural Out-of-Distribution Detectors. That paper uses norm balancing across classes at equal depth. We reuse that balancing so that radius means tree depth rather than class frequency.
Research Program
The front end is the Audity two-branch encoder. The outputs of its two branches are concatenated, passed through a small projection, and mapped onto the Lorentz hyperboloid with the exponential map at the origin.
We build a taxonomy from the published architecture metadata of MLAAD and ADD 2023 generators. It has three levels below a root: broad synthesis paradigm, acoustic-model family, and vocoder or waveform decoder. Bona fide speech is a separate branch. Closed-source commercial engines whose internals are unknown are attached only at the deepest level we can justify, and that uncertainty is part of the test.
Each taxonomy node $v$ gets a learnable prototype $p_v$ on the hyperboloid. Distance is
$$d(x,y)=\operatorname{arccosh}(-\langle x,y\rangle_L),$$
where $\langle x,y\rangle_L=-x_0y_0+\sum_i x_iy_i$ is the Lorentzian inner product.
Training combines four terms. First, a prototype loss pulls each clip embedding toward its leaf prototype and away from the others. Second, a pairwise contrastive term on clip embeddings makes same-generator pairs close. Third, an entailment-cone constraint, adapted from Hyperbolic Entailment Cones, requires each child prototype to sit inside its parent's cone. This encodes the asymmetric "this vocoder belongs to this family" relation that a symmetric triplet loss cannot. Fourth, a radial balancing term ties prototype distance from the origin to tree depth.
To reduce speaker leakage, a gradient-reversal speaker classifier is applied to the projection. Pairs with the same generator but different speakers are oversampled.
Inference produces two outputs. Verification scores a clip pair by $-d(x,y)$. Attribution descends the tree from the root and stops at the deepest node $v$ whose prototype is closer than all its siblings' by a margin calibrated on held-out known generators. The analyst therefore gets an answer like "GAN-vocoder family, specific system unknown" rather than a confident wrong leaf.
Baselines
The naive reference point is cosine similarity between mean-pooled frozen w2v-BERT 2.0 embeddings, with no forensic training. This tells us how much of verification comes from the backbone alone. No reported result was found on Lacuna for this baseline on MLAAD verification, so we will measure it.
The strongest existing method is Audity from Audio Deepfake Verification, which reports 10.02% verification EER and 89.87% accuracy on MLAAD. It also reports a drop from 25.0% to 4.2% EER on ADD 2023 when going from 1-shot to 5-shot enrollment. We will retrain Audity with the identical encoders, training split, held-out generator lists, augmentation, and hyperparameter search budget, so that differences come only from the head and losses. We will also reproduce its published protocol to confirm our retraining matches the reported numbers. Beating it means lower verification EER on unseen-generator pairs at equal or smaller embedding dimension, with non-overlapping confidence intervals across three seeds.
The ablation that isolates the key ingredient is a Euclidean twin of our model. It has the same taxonomy, the same prototype and contrastive losses, Euclidean order-embedding constraints in place of cones, and norm-to-depth regularization, all at matched dimension. A second ablation removes the taxonomy terms from the hyperbolic model, leaving a hyperbolic Audity. Together these two ablations separate the contribution of geometry from the contribution of hierarchical labels.
We will also adapt Poin-HierNet's latent prototype hierarchy to verification by scoring pairs with Poincaré distance. This tests whether a learned, unlabeled hierarchy suffices. No reported result was found on Lacuna for Poin-HierNet on verification or attribution.
Why not just…?
Why not just add hierarchical multi-task heads to Audity in Euclidean space? This is the real threat. A Euclidean model with one classifier per taxonomy level can also abstain level by level. The Euclidean twin ablation is exactly this comparison. If it matches us on partial attribution and on verification at all dimensions, we will report that the taxonomy, not the curvature, is the contribution. Our expectation, which rests on the exponential-volume argument and the WordNet results in the Lorentz paper, is that the hyperbolic advantage appears mainly at low dimension and on deep, wide subtrees. The dimension sweep is designed to show this, and the claim should be read as conditional on it.
Why not just increase Euclidean embedding dimension? At high dimension a flat space may hold a three-level tree of roughly a hundred leaves without much crowding. We will sweep dimension from 8 to 256 for both geometries. If the curves meet by 64 dimensions, the practical argument narrows to compact embeddings for large-scale forensic search, and we will state that.
Why not just use a frozen SSL backbone with a linear head, as in Towards generalisable and calibrated audio deepfake detection? That recipe reaches 8.8% average EER across eight out-of-domain binary-detection sets. However, a linear head produces no notion of source similarity or depth. The naive cosine baseline and a linear-probe-plus-cosine variant will show how far that recipe falls short on verification.
Proposed Experiments
We use MLAAD and ADD 2023 with three held-out protocols of increasing difficulty. In the first, the unseen generator reuses a training vocoder and acoustic family. In the second, the unseen generator has a new vocoder within a known family. In the third, a whole paradigm is held out. Commercial engines are reserved entirely for testing.
The metrics are as follows. We measure verification EER on unseen-source pairs. We measure depth-resolved attribution accuracy, meaning the fraction of clips whose predicted stopping node is a correct ancestor, together with the mean depth reached. We measure the rank correlation between embedding radius and the true depth of known ancestry for unseen generators. We measure EER on same-generator, different-speaker versus different-generator, same-speaker pairs to quantify speaker leakage. We also report training stability, as gradient norms and NaN rate, for Lorentz versus Poincaré versions of the same model.
The decisive check is the third protocol compared against the Euclidean twin at 16 and 32 dimensions. The hyperbolic model must stop at the correct internal node significantly more often and must not wrongly commit to leaves. It must also show positive correlation between radius and known depth. If it fails this check while still winning plain verification EER, the contribution reduces to compression, and we will say so.
As a sanity check that the forensic embedding has not damaged basic detection, we will report binary EER on In-the-Wild against Poin-HierNet's 4.91%. We do not claim to improve on that number.
References
- Audio Deepfake Verification
- Generalizable Audio Deepfake Detection via Hierarchical Structure Learning and Feature Whitening in Poincaré sphere
- HierCon: Hierarchical Contrastive Attention for Audio Deepfake Detection
- Hyperbolic Face Anti-Spoofing
- Hyp-OC: Hyperbolic One Class Classification for Face Anti-Spoofing
- Rethinking Generalizable Face Anti-Spoofing via Hierarchical Prototype-Guided Distribution Refinement in Hyperbolic Space
- Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry
- Balanced Hyperbolic Embeddings Are Natural Out-of-Distribution Detectors
- Hyperbolic Entailment Cones for Learning Hierarchical Embeddings
- Towards generalisable and calibrated audio deepfake detection with self-supervised representations