Traditional deepfake detection operates on a simple binary premise: determining whether a given audio clip is real or fake. While this approach can flag a deceptive clip, it leaves a critical forensic gap (Lu et al., 2024; Lin I, 2025). In legal proceedings, intellectual property disputes, or disinformation tracking, knowing that an audio file is synthetic is only half the battle; investigators need to know where it came from (Zhang et al., 2023; Xie et al., 2024).
To bridge this gap, researchers are shifting toward source tracing—the practice of identifying the specific generative algorithms, architectures, or model checkpoints responsible for synthesizing a speech waveform (Yi et al., 2024; Stan et al., 2025).
Architectural Fingerprints in the Waveform
How can an algorithm look at a seamless audio clip and pinpoint the exact model that generated it? The answer lies in "architectural fingerprints." Just as a physical firearm leaves distinct striations on a bullet, neural speech synthesis pipelines leave unique, indelible traces in the audio waveforms they produce (Zhang et al., 2023; Yi et al., 2022).
These fingerprints manifest at multiple levels of the synthesis process:
- Codec and Vocoder Primitives: Lin I demonstrates that fundamental engineering choices in neural audio codecs—such as the specific mathematical quantization method or whether the decoder operates in the time or frequency domain—imprint unique technical signatures on the output.
- Prosodic Biases: Generative models do not just struggle with low-level acoustic artifacts; they also possess systemic biases in how they model macro-level speech patterns. Phukan et al. (2025) show that models leave "prosodic signatures"—inherent rhythms, pitch variations, and intonation tendencies—that can be extracted using speech pre-trained models like Whisper.
Because these signatures depend on the underlying mathematics of the generator rather than the speaker's voice, source tracing systems can isolate the identity of the algorithm regardless of whose voice is being spoofed (Firc et al., 2025).
Component-Level Tracing for Generalizability
A naive approach to source tracing would treat every new deepfake model as a completely distinct class in a massive multi-class classifier. However, this approach fails when encountering a brand-new synthesis system.
To solve this, Klein et al. (2024) and Falez et al. (2025) propose decomposing the problem into component-level tracing. A typical speech synthesis pipeline consists of two main stages: an acoustic model (which maps text to an intermediate representation like a mel-spectrogram) and a vocoder or neural codec (which synthesizes the final waveform).
[Text Input] ──> [ Acoustic Model ] ──> [ Vocoder / Codec ] ──> [ Synthetic Waveform ]
│ │
▼ ▼
(Acoustic Fingerprint) (Codec/Vocoder Fingerprint)
│ │
└───────────┬────────────┘
▼
[Component-Level Detector]
When developers release a new deepfake tool, they rarely invent both components from scratch; instead, they often pair a novel acoustic model with a well-known vocoder, or vice versa. By training forensic models to independently recognize the fingerprints of specific acoustic models and vocoders, the system becomes highly modular (Hong et al., 2025). If an attacker deploys a novel hybrid system, a component-level detector can still generalize, identifying the known building blocks used to construct it (Klein et al., 2024).
Navigating the Open-Set Wilderness
In real-world forensics, detectors operate in an "open-set" environment where they will inevitably encounter audio generated by completely unknown, out-of-distribution (OOD) models (Klein et al., 2025).
To handle this, researchers leverage metric learning—borrowing techniques originally designed for speaker verification (Koutsianos et al., 2025). Instead of forcing a model to pick from a fixed menu of known algorithms, metric learning trains neural networks to project audio clips into a continuous embedding space. In this space, clips generated by the same system cluster together, while clips from different systems are pushed apart (Müller et al., 2022).
When a novel deepfake algorithm appears, it maps to an entirely new, isolated cluster in this embedding space. Researchers can then use metrics like "softmax energy" to flag these clips as out-of-distribution (Klein et al., 2025), or apply dispersion strategies that deliberately separate unknown fake distributions from known real ones (Xie et al., 2024). This allows forensic tools to determine with confidence when a clip belongs to a brand-new family of generators, even enabling training-free attribution down to specific model checkpoints (Stan et al., 2025).
Key Foundational Work
An Initial Investigation for Detecting Vocoder Fingerprints of Fake Audio (2022) — contributed 3 concept elements to this direction.
Attacker Attribution of Audio Deepfakes (2022) — contributed 2 concept elements to this direction.
Distinguishing Neural Speech Synthesis Models Through Fingerprints in Speech Waveforms (2023) — contributed 1 concept elements to this direction.