Speech Synthesis and Audio Deepfake Detection — Literature Catalog

S01 · Spark-TTS

Spark-TTS: An Efficient LLM-Based Text-to-Speech Model with Single-Stream Decoupled Speech Tokens

Xinsheng Wang et al. · first public 2025-03-03 · checked 2503.01710v1 · updated 2025-03-03

Question. Separate speaker identity from changing speech content.

Approach. A global speaker representation conditions reconstruction; a single token stream carries changing speech information.

Evidence boundary. A useful 2025 anchor; separation alone does not establish a new contribution.

S02 · IndexTTS2

IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech

Siyi Zhou et al. · first public 2025-06-23 · checked 2506.21619v2 · updated 2025-09-03

Question. Control emotion and duration without losing the requested voice.

Approach. Separate emotion and identity prompts; add token-count duration control and text-based emotional instructions.

Evidence boundary. Judge duration, intelligibility, emotion and speaker identity separately.

S03 · DisCo-Speech

DisCo-Speech: Controllable Zero-Shot Speech Generation with A Disentangled Speech Codec

Tao Li et al. · first public 2025-12-15 · checked 2512.13251v3 · updated 2026-01-04

Question. Transfer style without copying the reference speaker.

Approach. A disentangled codec separates content, prosody and timbre for controlled zero-shot generation.

Evidence boundary. Independent control needs tests on previously unseen combinations.

S04 · Fish Audio S2

Fish Audio S2 Technical Report

Shijia Liao et al. · first public 2026-03-09 · checked 2603.08823v2 · updated 2026-03-11

Question. Follow expressive instructions in speech generation.

Approach. A dual autoregressive architecture supports expressive, multi-speaker and multi-turn synthesis.

Evidence boundary. Treat report comparisons as protocol-specific evidence.

S05 · FC-TTS

FC-TTS: Style and Timbre Control in Zero-Shot Text-to-Speech with Disentangled Speech Representations

Yoonhyung Lee et al. · first public 2026-05-23 · checked 2605.24618v1 · updated 2026-05-23

Question. Choose style and voice from different recordings.

Approach. Separate reference representations and consistency objectives support independent style and timbre control.

Evidence boundary. A changed style can affect perceived speaker identity; evaluate both together.

S06 · FireRedAudio

FireRedAudio: A General-Purpose Audio Language Model with Decoupled Continuous Representations for Understanding and Generation

Feiyu Shen et al. · first public 2026-08-25 · checked 2608.24168v2 · updated 2026-08-27

Question. Support synthesis and speech editing in a unified audio model.

Approach. Use separate continuous input paths for understanding and reconstructible generation; condition a flow model for speech.

Evidence boundary. Included for synthesis and editing; its broader understanding tasks are outside this map.

S07 · Fish Audio Drama 3

Fish Audio Drama 3 Technical Report

Dawei Han et al. · first public 2026-10-13 · checked 2610.18010v2 · updated 2026-10-14

Question. Generate and revise expressive multi-speaker scenes.

Approach. Scale speech modeling, predict multiple codebooks and use human preference training.

Evidence boundary. The evaluation includes regeneration of failed outputs; single-attempt reliability needs separate testing.

S08 · Kanade

Kanade: A Simple Disentangled Tokenizer for Spoken Language Modeling

Zhijie Huang et al. · first public 2026-01-31 · checked 2602.00594v2 · updated 2026-05-05

Question. Keep pronunciation and rhythm while separating stable acoustic information.

Approach. A single discrete stream carries changing speech features; a global continuous vector carries acoustic constants.

Evidence boundary. Full-text caveat: the presented tokenizer is not natively streaming.

S09 · SiTok

Scaling Speech Tokenizers with Diffusion Autoencoders

Yuancheng Wang et al. · first public 2026-02-06 · checked 2602.06602v1 · updated 2026-02-06

Question. Compress speech without sacrificing useful generation information.

Approach. Scale a diffusion-autoencoder tokenizer with supervised semantic information and low-rate speech tokens.

Evidence boundary. Reconstruction quality alone does not measure downstream generation quality.

S10 · XY-Tokenizer

XY-Tokenizer: Mitigating the Semantic-Acoustic Conflict in Low-Bitrate Speech Codecs

Yitian Gong et al. · first public 2026-07 · ACL 2026

Question. Balance linguistic information and acoustic detail at low bitrate.

Approach. Use staged multitask training to align discrete speech features with text while preserving reconstruction.

Evidence boundary. Evaluate representation quality using downstream understanding and generation, not bitrate alone.

S11 · ReLMCodec

ReLMCodec: Designing Predictable Speech Tokens from Pre-Quantization Phoneme Structure

Zixiang Wan et al. · first public 2026-08-08 · checked 2608.08286v1 · updated 2026-08-08

Question. Make speech tokens easier for a language model to predict.

Approach. Organize pre-quantization features using phoneme structure to improve token predictability.

Evidence boundary. Compare language-model difficulty alongside speech reconstruction quality.

S12 · Content representation study

A Comprehensive Study of Content Representations for Speech Synthesis

Diego Torres et al. · first public 2026-09-25 · checked 2609.30975v1 · updated 2026-09-25

Question. Measure what content representations actually retain.

Approach. Compare representations under one generative framework along content, speaker and prosody dimensions.

Evidence boundary. The study links disentanglement to both objective and information capacity.

S13 · Q-SPT

Q-SPT: Learnable Query-Based Compression for Low-Frame-Rate Speech Tokenization

Jeeyoung Yun et al. · first public 2026-10-01 · checked 2610.01492v1 · updated 2026-10-01

Question. Reduce speech token rate while keeping useful detail.

Approach. Learn query-based compression for semantic and acoustic features.

Evidence boundary. A lower frame rate needs controlled quality and downstream-task comparisons.

S14 · Where Does Timbre Hide?

Where Does Timbre Hide? Localizing Speaker Information in Speech Signals

Yiquan Zhou et al. · first public 2026-10-21 · checked 2610.29150v1 · updated 2026-10-21

Question. Locate residual speaker information in signals and tokens.

Approach. Probe frequency bands, signal components, model layers and tokenizers using speaker identity and listener tests.

Evidence boundary. Speaker identity is a measurable proxy for timbre, not its complete definition.

S15 · CosyVoice 3

CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training

Zhihao Du et al. · first public 2025-05-23 · checked 2505.17589v2 · updated 2025-05-27

Question. Improve synthesis on difficult real-world text and references.

Approach. Combine scale, a multitask speech tokenizer and post-training feedback.

Evidence boundary. Useful system-level evidence; isolate the contribution of data, objectives and model scale.

S16 · Qwen3-TTS

Qwen3-TTS Technical Report

Hangrui Hu et al. · first public 2026-01-22 · checked 2601.15621v1 · updated 2026-01-22

Question. Support controllable multilingual synthesis with efficient speech generation.

Approach. Present alternative tokenization and generation tracks, including causal speech decoding.

Evidence boundary. Token rate, codebook count and end-to-end latency are different quantities.

S17 · VoxCPM2

VoxCPM2 Technical Report

Yixuan Zhou et al. · first public 2026-06-05 · checked 2606.06928v1 · updated 2026-06-05

Question. Model speech directly in continuous acoustic representations.

Approach. Predict acoustic patches autoregressively and generate local detail with diffusion.

Evidence boundary. It avoids an external discrete token stream; internal quantization still exists.

S18 · End-to-end discrete-token TTS

End-to-End Training for Discrete Token LLM based TTS System

Changfeng Gao et al. · first public 2026-06-08 · checked 2606.09234v1 · updated 2026-06-08

Question. Reduce mismatch between separately trained synthesis components.

Approach. Train tokenizer representations with reconstruction, prediction and recognition objectives; then use downstream feedback for the language model.

Evidence boundary. This is staged joint optimization; not every parameter is updated in every stage.

S19 · Qwen-Audio-3.0-TTS

Qwen-Audio-3.0-TTS: Freely Controllable and Highly Robust Speech Synthesis with Multi-Stage Training Paradigm

Bajian Xiang et al. · first public 2026-07-27 · checked 2607.23938v1 · updated 2026-07-27

Question. Improve control and robustness through coordinated training.

Approach. Use a multi-stage recipe that combines language-model training, flow-model training and reinforcement learning.

Evidence boundary. Separate improvements in pronunciation, control, reference robustness and long-text behavior.

S20 · Luna-TTS

Luna-TTS Family Technical Report

Feng Yin et al. · first public 2026-08-12 · checked 2608.11593v1 · updated 2026-08-12

Question. Reduce sequential decoding while retaining controllable generation.

Approach. Denoise a discrete token grid in parallel; use blockwise generation for the realtime variant.

Evidence boundary. The realtime variant uses 1.28-second blocks; reported warm latency is protocol-specific.

S21 · The Evolving Bottleneck

The Evolving Bottleneck in Speech Generation: Interface Co-design and Staged Alignment from CosyVoice to Qwen-Audio-3.0-TTS

Qian Chen et al. · first public 2026-09-15 · checked 2609.16514v1 · updated 2026-09-15

Question. Explain why improving one synthesis component can stop helping.

Approach. A technical analysis traces interface design and staged alignment from CosyVoice to Qwen-Audio-3.0-TTS.

Evidence boundary. Background synthesis of one model lineage; not an independent new benchmark or universal proof.

S22 · LoCA

LoCA: Local Credit Assignment for Reinforcement Learning of Autoregressive Flow-Matching TTS

Jie Liu et al. · first public 2026-10-20 · checked 2610.28035v1 · updated 2026-10-20

Question. Give local feedback when only part of an utterance is wrong.

Approach. Use ASR alignment to redistribute rewards across acoustic frames and constrain positive and negative regression updates.

Evidence boundary. Experiments use VoxCPM2; transfer to other synthesis architectures remains a research question.

S23 · MegaTTS 3

MegaTTS 3: Sparse Alignment Enhanced Latent Diffusion Transformer for Zero-Shot Speech Synthesis

Ziyue Jiang et al. · first public 2025-02-26 · checked 2502.18924v4 · updated 2025-03-28

Question. Balance reliable text alignment with natural speech.

Approach. Sparse alignment guides latent diffusion; additional conditioning controls accent intensity.

Evidence boundary. Use difficult sentences to test alignment, rather than relying only on average naturalness.

S24 · ZipVoice

ZipVoice: Fast and High-Quality Zero-Shot Text-to-Speech with Flow Matching

Han Zhu et al. · first public 2025-06-16 · checked 2506.13053v3 · updated 2025-08-07

Question. Make high-quality zero-shot synthesis cheaper to run.

Approach. Use compact Zipformer networks, initial text alignment and flow distillation to reduce inference cost.

Evidence boundary. Speed comparisons depend on model, hardware, sampling settings and quality targets.

S25 · ZipVoice-Dialog

ZipVoice-Dialog: Non-Autoregressive Spoken Dialogue Generation with Flow Matching

Han Zhu et al. · first public 2025-07-12 · checked 2507.09318v2 · updated 2026-04-14

Question. Synthesize scripted dialogue with correct speaker turns.

Approach. Use alignment curriculum and speaker-turn embeddings in non-autoregressive flow matching; release OpenDialog.

Evidence boundary. Scripted multi-speaker synthesis does not establish interactive full-duplex dialogue.

S26 · VibeVoice

VibeVoice Technical Report

Zhiliang Peng et al. · first public 2025-08-26 · checked 2508.19205v1 · updated 2025-08-26

Question. Maintain voices and context across long multi-speaker recordings.

Approach. Combine low-rate continuous representations with autoregressive latent generation and a diffusion head.

Evidence boundary. Long generation capability does not itself prove consistent quality for the whole recording.

S27 · VoXtream

VoXtream: Full-Stream Text-to-Speech with Extremely Low Latency

Nikita Torgashov et al. · first public 2025-09-19 · checked 2509.15969v2 · updated 2026-01-26

Question. Start speaking before the full input text is available.

Approach. Use incremental text processing and bounded future context in a full-stream TTS pipeline.

Evidence boundary. Measure quality against available text lookahead and first-audio latency.

S28 · VoXtream2

VoXtream2: Full-stream TTS with dynamic speaking rate control

Nikita Torgashov et al. · first public 2026-03-13 · checked 2603.13518v1 · updated 2026-03-13

Question. Control speaking rate during full-stream synthesis.

Approach. Extend full-stream generation with dynamic rate control and prompt masking during training.

Evidence boundary. Test rapidly changing text and rate requests as well as fixed sentences.

S29 · MOSS-TTS

MOSS-TTS Technical Report

Yitian Gong et al. · first public 2026-03-18 · checked 2603.18090v2 · updated 2026-03-20

Question. Scale discrete-token speech generation for complex long contexts.

Approach. A speech language-model system targets multilingual, controllable and long-form generation.

Evidence boundary. Assess content errors, voice consistency and long-context stability separately.

S30 · SwanBench-Speech

Comprehensive Benchmarking of Long-Form Speech Generation in Diverse Scenarios

Changhao Pan et al. · first public 2026-05-27 · checked 2605.28618v1 · updated 2026-05-27

Question. Evaluate long-form quality beyond short-utterance scores.

Approach. Benchmark diverse scenarios across acoustics, semantics and expressiveness using seven metrics.

Evidence boundary. Automated scores require human validation for discourse-level expression and coherence.

S31 · TontaubeV1

TontaubeV1: Streaming Text-to-Speech with Hierarchical Codec Modeling and Bounded Context

Fritz Cremer et al. · first public 2026-09-08 · checked 2609.08703v2 · updated 2026-10-13

Question. Stream long speech without unbounded context growth.

Approach. Use hierarchical codec modeling and bounded context for streaming synthesis.

Evidence boundary. Streaming audio output alone does not prove support for incremental text input.

S32 · StreamAlign

StreamAlign: Streaming Text-Aligned Speech Tokenization

Kang-wook Kim et al. · first public 2026-09-09 · checked 2609.09719v1 · updated 2026-09-09

Question. Build speech tokens without waiting for the complete recording.

Approach. Combine online alignment, ASR guidance and word-boundary prediction for streaming text-aligned tokens.

Evidence boundary. A representation study with speech-continuation evidence; not a complete TTS deployment claim.

D01 · ASVspoof 5 resources

ASVspoof 5: Design, Collection and Validation of Resources for Spoofing, Deepfake, and Adversarial Attack Detection Using Crowdsourced Speech

Xin Wang et al. · first public 2025-02-13 · checked 2502.08857v4 · updated 2025-04-24

Question. Build a demanding shared test of spoofing and deepfake detection.

Approach. Release and validate crowdsourced speech resources with diverse synthesis and adversarial conditions.

Evidence boundary. A resource paper; pair it with the later evaluation report instead of counting two independent benchmarks.

D02 · AUDDT

AUDDT: A Unified Benchmark Toolkit for Audio and Speech Deepfake Detectors

Yi Zhu et al. · first public 2025-09-25 · checked 2509.21597v2 · updated 2026-06-03

Question. Evaluate the same detector across many existing datasets.

Approach. Provide a unified toolkit and metadata analysis covering 31 datasets in the checked revision.

Evidence boundary. A broad toolkit does not remove overlap, label or recording-condition biases in its source datasets.

D03 · TwinShift

TwinShift: Benchmarking Audio Deepfake Detection across Synthesizer and Speaker Shifts

Jiyoung Hong et al. · first public 2025-10-27 · checked 2510.23096v1 · updated 2025-10-27

Question. Separate unseen-speaker effects from unseen-generator effects.

Approach. Create evaluation splits with disjoint speakers and synthesis systems.

Evidence boundary. Use the split design to test transfer; avoid treating one aggregate score as universal generalization.

D04 · ASVspoof 5 evaluation

ASVspoof 5: Evaluation of Spoofing, Deepfake, and Adversarial Attack Detection Using Crowdsourced Speech

Xin Wang et al. · first public 2026-01-07 · checked 2601.03944v3 · updated 2026-04-11

Question. Measure detector performance on the shared challenge.

Approach. Analyze submitted systems, neural-codec attacks, adversarial conditions and evaluation behavior.

Evidence boundary. Challenge rankings are tied to its data and operating conditions.

D05 · Spoof-SUPERB

A SUPERB-Style Benchmark of Self-Supervised Speech Models for Audio Deepfake Detection

Hashim Ali et al. · first public 2026-03-02 · checked 2603.01482v1 · updated 2026-03-02

Question. Compare pretrained speech representations for detection fairly.

Approach. Evaluate 20 self-supervised models across in-domain, out-of-domain and degraded audio.

Evidence boundary. Backbone rankings depend on training and evaluation choices.

D06 · LRLspoof

When Spoof Detectors Travel: Evaluation Across 66 Languages in the Low-Resource Language Spoofing Corpus

Kirill Borodin et al. · first public 2026-03-02 · checked 2603.02364v4 · updated 2026-06-27

Question. Test whether a detector transfers across languages.

Approach. Evaluate synthetic speech from 24 TTS systems in 66 languages using transferred decision thresholds.

Evidence boundary. Spoof rejection rates without target genuine speech do not establish target-language false-alarm rates.

D07 · Generalizable Detection

Generalizable Detection of Audio Deepfakes

Jose A. Lopez et al. · first public 2025-07-02 · checked 2507.01750v1 · updated 2025-07-02

Question. Identify training choices that improve unseen-condition detection.

Approach. Compare pretrained speech backbones, augmentation and loss choices across out-of-domain evaluations.

Evidence boundary. Use held-out generator and dataset tests to distinguish transfer from dataset fit.

D08 · Continual detection with UAP

Continual Audio Deepfake Detection via Universal Adversarial Perturbation

Wangjie Li et al. · first public 2025-11-25 · checked 2511.19974v1 · updated 2025-11-25

Question. Learn new attacks without forgetting earlier fake speech.

Approach. Use universal adversarial perturbations to retain information about past distributions during continual training.

Evidence boundary. The task sequence and replay assumptions define the strength of the result.

D09 · Advanced TTS evaluation

Audio Deepfake Detection in the Age of Advanced Text-to-Speech models

Robin Singh et al. · first public 2026-01-28 · checked 2601.20510v1 · updated 2026-01-28

Question. Check whether newer synthesis families break existing detectors.

Approach. Generate a controlled corpus with three synthesis systems and compare four detection approaches.

Evidence boundary. Its multi-view findings concern the tested systems, not all modern TTS.

D10 · CoReLoop

CoReLoop: Parameter-Efficient Controlled Recurrent Refinement for Audio Deepfake Detection

Kunyu Feng et al. · first public 2026-09-17 · checked 2609.19818v3 · updated 2026-10-13

Question. Improve an existing detector without retraining its whole backbone.

Approach. Add controlled recurrent refinement through low-rank adaptation and lightweight modules.

Evidence boundary. Extra inference passes have a compute cost; compare quality at matched deployment budgets.

D11 · RF-Prompt

Learning as Deepfakes Evolve: RF-Prompt for Continual Audio Deepfake Detection

Yuankun Xie et al. · first public 2026-09-29 · checked 2609.37586v1 · updated 2026-09-29

Question. Separate learning new generators from learning new real-speech domains.

Approach. Introduce a controlled incremental protocol and shared real-speech prompts with expanding fake experts.

Evidence boundary. The protocol controls specific shifts; deployment still needs broader chronological testing.

D12 · Task-aware compression

Task-Aware Joint Pruning and Distillation for Efficient Audio Deepfake Detection

Miao He et al. · first public 2026-10-04 · checked 2610.05264v1 · updated 2026-10-04

Question. Put large pretrained detectors on smaller devices.

Approach. Jointly prune model structure and distill detection knowledge across domains.

Evidence boundary. Compression must preserve unseen-condition detection and calibration, not only model size.

D13 · Adaptive compression and perturbation

Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection

Jianqiao Cui et al. · first public 2026-10-12 · checked 2610.16049v1 · updated 2026-10-12

Question. Improve compact representations for varying spoof artifacts.

Approach. Combine adaptive feature compression with targeted mel-band regularization.

Evidence boundary. Reported comparisons cover selected datasets and settings; no universal winner follows.

D14 · PartialEdit

PartialEdit: Identifying Partial Deepfakes in the Era of Neural Speech Editing

You Zhang et al. · first public 2025-08 · Interspeech 2025

Question. Detect seamless edits produced by neural speech-editing systems.

Approach. Build an edited-speech dataset and evaluate detection and temporal localization.

Evidence boundary. The paper reports failure of detectors transferred from PartialSpoof; realistic editing changes the test distribution.

D15 · Split and Conquer

Split and Conquer Partial Deepfake Speech

Inbal Rimon et al. · first public 2026-04-03 · checked 2604.02913v1 · updated 2026-04-03

Question. Locate short fake regions inside otherwise genuine speech.

Approach. First detect boundaries, then classify segments; train with multiple segment lengths.

Evidence boundary. Validate localization and classification separately on edits from unseen systems.

D16 · Half-Truth / CAFNet

Half-Truth Audio Detection and Localisation: A Lightweight Cross-Attentive Architecture and a Cross-Corpus Diagnostic Study

S. Sutharya et al. · first public 2026-05-28 · checked 2605.29531v3 · updated 2026-09-02

Question. Detect and localize edits with a small model.

Approach. Fuse complementary acoustic features using cross-attention and predict fake-region boundaries.

Evidence boundary. The checked revision reports near-chance binary transfer to PartialSpoof despite stronger in-domain results.

D17 · What Do Detectors Hear?

What Do Deepfake Speech Detectors Actually Hear?

Vojtěch Staněk et al. · first public 2026-06-09 · checked 2606.10912v1 · updated 2026-06-09

Question. Identify the acoustic evidence behind detector decisions.

Approach. Use time-aligned attribution, manual cue annotation and causal masking with three WavLM-based detectors.

Evidence boundary. Attribution explains these models on the tested data; it is not forensic proof for every recording.

D18 · Zero-shot conversation localization

Zero-Shot Temporal Localisation of Audio Deepfakes in Multi-Speaker Conversations

Soumyadeep Roy · first public 2026-09-09 · checked 2609.10051v1 · updated 2026-09-09

Question. Locate inserted synthetic sentences in multi-speaker recordings.

Approach. Wrap frozen detectors with window scoring and temporal decoding without retraining.

Evidence boundary. The benchmark uses constructed conversations; natural recordings need separate validation.

D19 · Audio watermarking SoK

SoK: How Robust is Audio Watermarking in Generative AI models?

Yizhu Wen et al. · first public 2025-03-24 · checked 2503.19176v2 · updated 2025-03-27

Question. Assess whether embedded provenance survives signal changes.

Approach. Systematically examine watermark robustness in generative-audio settings.

Evidence boundary. Adjacent provenance evidence; absence of a detectable watermark is not proof of genuine speech.

D20 · RADAR 2026

RADAR Challenge 2026: Robust Audio Deepfake Recognition under Media Transformations

Hieu-Thi Luong et al. · first public 2026-05-10 · checked 2605.09568v3 · updated 2026-05-25

Question. Evaluate detection after ordinary media processing.

Approach. Benchmark compressed, resampled, noisy and reverberant audio under shared challenge conditions.

Evidence boundary. Report threshold behavior and false alarms alongside benchmark error rates.

D21 · MixFake

MixFake: Benchmarking and Enhancing Audio Deepfake Detection in Diverse Real-world Mixed Audio

Qingcao Li et al. · first public 2026-05-22 · checked 2605.23201v2 · updated 2026-10-08

Question. Detect synthetic components in mixed real-world audio.

Approach. Create mixed-authenticity foreground and background examples; inject signal-level priors through multi-stream prompts.

Evidence boundary. Mixed audio needs source-aware labels; a single file label can hide the target task.

D22 · Proteus

Proteus: Automated Adversarial Robustness Testing for Audio Deepfake Detectors

Nicolas M. Müller et al. · first public 2026-06-28 · checked 2606.29544v1 · updated 2026-06-28

Question. Find failures caused by combinations of common transformations.

Approach. Automate robustness testing against a production detector and use failures for targeted retraining.

Evidence boundary. Evidence comes from a specific detector; other systems need independent audits.

D23 · AffectDF

AffectDF: The Most Comprehensive Benchmark for Speech Deepfake Detection against Emotionally Expressive Attacks

Aurosweta Mahapatra et al. · first public 2026-08-06 · checked 2608.05507v1 · updated 2026-08-06

Question. Test detectors against expressive rather than neutral fake voices.

Approach. Evaluate emotional TTS, conversion and audio-language-model attacks across emotion conditions.

Evidence boundary. Adding emotional training data does not consistently solve transfer to unseen attacks.

D24 · AT-ADD

AT-ADD: A Benchmark and Challenge for Robust and All-Type Audio Deepfake Detection

Yuankun Xie et al. · first public 2026-08-24 · checked 2608.23437v2 · updated 2026-10-15

Question. Measure robustness to unseen generators, recording channels and replay.

Approach. Run robust-speech and all-audio tracks with shared baselines and submitted systems.

Evidence boundary. This map uses the speech track; the all-audio track also includes music, singing and environmental sounds.

D25 · Watermark impact on detectors

On the Robustness of Audio Deepfake Detection under Audio Watermarking

Zi Qian Yong et al. · first public 2026-08-25 · checked 2608.24159v1 · updated 2026-08-25

Question. Check whether benign provenance marks disturb detection.

Approach. Measure detector scores and representation shifts after watermarking across several datasets.

Evidence boundary. The impact is dataset-dependent; watermarking and blind detection need joint testing.

D26 · GenTraceBench

GenTraceBench: A Benchmark for Tracing Audio Deepfakes Across Pre- and Post-training Stages

Li Wang et al. · first public 2026-09-18 · checked 2609.21738v1 · updated 2026-09-18

Question. Test whether generator fingerprints survive model adaptation.

Approach. Hold texts and prompts fixed across pretraining and post-training variants; test detection and source tracing.

Evidence boundary. Fingerprint stability differs by adaptation method and forensic backbone.

D27 · Adaptive purification

Detection-Guided Adaptive Purification with Diffusion Models for Robust Audio Deepfake Detection

Muhammed Salih Kayhan et al. · first public 2026-10-07 · checked 2610.10752v1 · updated 2026-10-07

Question. Preserve detection evidence while removing adversarial perturbations.

Approach. Adjust diffusion purification strength per input using detector score changes.

Evidence boundary. The tested attack and detector settings define the defense claim; benign-audio quality also matters.