arXiv_SD
-
Beyond Prompt Adherence: Auditing Attribute-Level Voice Control in Speech Generation
Abstract
Natural-language descriptions have become a flexible interface for controlling generated speech. Existing evaluations largely assess whether an output matches a prompt, but prompt matching alone does not reveal whether characteristics outside the intended change remain stable. We examine this distinction through a controlled paired audit of three speech-generation systems: CosyVoice3, VoxCPM2, and Fish-Speech-S2. The evaluation contains 5,940 outputs spanning six reference speakers, ten texts, three random seeds, and eleven conditions. Using acoustic, prosodic, content, and speaker measurements, we find that responses in the expected target direction are frequently accompanied by changes outside descriptor-specific signal-level target sets. This pattern remains among outputs whose target response exceeds baseline seed variation, and the accompanying changes differ substantially across systems. We further introduce VoDER-Cal, a training-free candidate selector that retains sufficiently strong target responses while favoring smaller off-target deviations. A three-candidate pool raises the joint success rate from 4.8% under single-sample direct generation to approximately 14% for all candidate-selection policies. Within the matched three-candidate budget, VoDER-Cal reduces held-out off-target deviation from 0.344 under target-only selection to 0.276 and improves listener-rated preservation. Preservation-sensitive evaluation therefore complements prompt-adherence evaluation, while candidate reranking offers a practical inference-time improvement. Code, configuration files, and analysis scripts are available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2608.00545
PDF
-
Hidden-Domain Routing for All-Type Audio Deepfake Detection
2026-08-01 07:32:54Yifan Gao, Yao Tian, Hongbin Suo, Haonan LuAbstract
All-type audio deepfake detection requires authenticity decisions across speech, environmental sound, singing voice, and music, while the audio type is unavailable at inference time. In AT-ADD Track2, this setting creates a hidden audio-domain condition: the binary real/fake label is shared across domains, but representation structure and detector-score behavior vary with audio type. We present a closed-condition routed system that first recovers the hidden audio domain and then interprets detector scores within the selected branch. The AudioType-BEATs-6s Router estimates audio type from a 6-second window; speech inputs are handled by the Speech-XLSR Expert, while sound, singing, and music rely on EAT-based general-audio experts with branch-local score interpretation. Development-set representation analysis, router-family comparisons, and component results show audio-domain separation and complementary detector strengths across audio types. On the official AT-ADD Track2 final evaluation, the system achieves 96.10% Track2 Macro-F1 and ranks first on the final leaderboard, with type-wise Macro-F1 scores of 88.07%, 98.18%, 99.07%, and 99.08% for speech, sound, singing, and music, respectively. These results support recovering the hidden audio domain before interpreting detector scores in all-type audio deepfake detection.
Abstract (translated)
URL
https://arxiv.org/abs/2608.00493
PDF
-
Normal-Anchored First-Order Model-Agnostic Meta-Learning based Whisper Fine-Tuning for Enhancing Fairness of Cleft Lip and Palate Speech Recognition
2026-07-31 18:08:53Susmita Bhattacharjee, Jagabandhu Mishra, H. S. Shekhawat, Ravi Jasuja, S. R. Mahadeva PrasannaAbstract
Automatic speech recognition (ASR) for cleft lip and palate (CLP) speech is difficult because acoustic and articulatory patterns vary across severity levels. This variability reduces the performance of pretrained ASR systems, and conventional fine-tuning may not generalize well under low-resource, heterogeneous CLP conditions. This work proposes Normal-Anchored First-Order Model-Agnostic Meta-Learning (NA-FOMAML) for adapting Whisper to CLP speech. The method uses a first-order bilevel meta-learning framework in which normal speech is used in the inner loop as a stable support condition, while CLP severity groups are used in the outer loop to improve post-adaptation robustness. This design aims to reduce the performance gap between normal and pathological speech. Experiments are conducted on the NMCPC and AIISH datasets using four normal-anchored training configurations. Frozen encoder, full encoder, and selected Whisper encoder-layer tuning strategies are evaluated, including layers 0--5, 4--11, 6--11, and 8--11, with decoder and projection-head adaptation. Results show that outer-loop training with only normal speech is insufficient. For NMCPC, full encoder tuning with Normal to Normal+Mild+Moderate gives WERs of 4.40%, 5.53%, 16.14%, and 52.07% for normal, mild, moderate, and severe speech. For AIISH, full encoder tuning with Normal to Normal+Mild+Moderate+Severe gives WERs of 2.48%, 19.66%, 14.05%, and 57.50%. A transcription-based phoneme-category analysis shows that severe CLP speech has high error rates across fricatives, affricates, nasals, liquids, plosives, and vowels. Overall, NA-FOMAML improves cross-severity robustness, but severe speech still requires severity-aware sampling, phoneme-aware loss functions, and augmentation targeting pressure consonant and resonance-related distortions.
Abstract (translated)
URL
https://arxiv.org/abs/2608.00186
PDF
-
ParaASR: Multi-Token Prediction for Fast and Long-Context LLM-Based Speech Recognition
2026-07-31 10:50:14Qingjian Lin, Yuxin Li, Haoyang Zhang, Jun Chen, Yechang Huang, Feng Tian, Xie Li, Xiangyu Tony Zhang, Daijiao Liu, Yuxin Zhang, Jinglan Gong, Bo Zhao, Fei Tian, Xuerui Yang, Gang Yu, Xiangyu Zhang, Daxin JiangAbstract
Audio-encoder-LLM-decoder architectures have become the dominant paradigm for modern automatic speech recognition (ASR), improving transcription quality through large-scale language modeling. However, the cost of autoregressive decoding scales with decoder size, creating a fundamental trade-off between recognition quality and serving latency. We argue this trade-off is not inherent: unlike open-ended text generation, ASR outputs are strongly anchored to the input speech signal, providing a natural inductive bias toward high-parallelism decoding. Building on this, we introduce ParaASR, an ASR system that leverages Multi-Token Prediction (MTP) to let a 4B LLM decoder emit multiple tokens per forward step. Starting from a publicly available audio-language foundation, the model first establishes a robust autoregressive recognizer and then aligns five future-token branches through a staged optimization recipe. At inference, it proposes a six-token continuation per step and admits only the verified prefix into the transcript, preserving the safety of standard autoregressive decoding. The average accepted length reaches 5.0 out of 6 proposed tokens, confirming that the deterministic structure of speech makes ASR an especially natural setting for multi-token decoding. ParaASR further retains a native 32K-context window and transcribes up to 30 minutes of audio in a single pass. Across diverse benchmarks, it attains average error rates of 2.97%, 3.68%, and 3.70% on Chinese, English, and long-form evaluations, respectively, while reaching a real-time factor (RTF) as low as 0.0053. These results show that decoder scaling, low-latency inference, and long-context transcription need not be competing goals when future-token proposals are anchored by the acoustic signal and guarded by autoregressive verification.
Abstract (translated)
URL
https://arxiv.org/abs/2607.29279
PDF
-
Do Music Foundation Models Embed Pitch in Helical Structure?
2026-07-31 07:13:54Hayato Yagi, Shinnosuke Takamichi, Rin Sato, Keitaro Tanaka, Shigeo MorishimaAbstract
This study analyzes the intermediate representations of music foundation models (MFMs) and reports the geometric structures used to represent pitch information. By inputting isolated musical notes into trained MFMs and analyzing their principal components, we reveal that the representations form a helical structure reflecting the octave periodicity of pitch. Furthermore, we show that the clarity and geometry of this helical structure vary not only across models but also with the acoustic properties of the input signals. Our analysis provides a novel approach for clarifying the internal mechanisms of MFMs.
Abstract (translated)
URL
https://arxiv.org/abs/2607.29086
PDF
-
Learning to Predict Performance-induced Emotion Differences in Classical Piano Music
Abstract
Music is often used as a medium for communicating emotion, with performers shaping perceived affect through interpretation. This study addresses the challenge of identifying and predicting subtle changes in perceived emotion that are exclusively due to differences in performance. We focus on classical solo piano music, using a set of 6 commercial recordings of Bach's Well-Tempered Clavier Book I, annotated in terms of valence and arousal. By encoding the recordings through performance-specific features only, we isolate performance information from aspects of the composition itself, which tend to dominate the overall perceived emotional category. A preliminary analysis validates that these features vary meaningfully across performers. We then propose a relative regression framework, Delta-VA, to predict deviations in valence-arousal relative to an ``average'' performance, thereby focusing on the changes in emotion brought about by a specific way of playing a piece. In addition to the standard $R^2$ regression score, we introduce geometric evaluation metrics to assess the preservation of pairwise differences between performances. Results indicate high directional consistency with the ground truth, but also a compression in prediction magnitude, indicating that the model tends to underestimate expressive performance effects.
Abstract (translated)
URL
https://arxiv.org/abs/2607.28876
PDF
-
Teffic-Audio: Tell Fact from Fiction
2026-07-30 15:21:38Wan Lin, Li Wang, Jindong Wang, Kunyu Feng, Zhizheng WuAbstract
Speech deepfake detection has expanded in scope with increasingly heterogeneous spoofing mechanisms, including speech synthesis, voice conversion, vocoder reconstruction, and neural-codec resynthesis. The resulting spoofing artifacts can be further shaped by variability in source speech, recording environments, and transmission channels. This variability makes robust generalization across heterogeneous conditions a central requirement for practical detection systems. This report presents Teffic-Audio, a general speech deepfake detection system designed for comprehensive evaluation environment. Teffic-Audio adopts a straightforward detector architecture consisting of a Conformer-based speech encoder, multi-head attentive statistics pooling, and a binary classifier. Rather than relying on additional architectural complexity, the system improves generalization through its training recipe, which integrates multi-source data, attack- and source-balanced sampling, and diverse audio augmentation. Trained only with open-source data, Teffic-Audio achieves a pooled EER of 1.454% on the 14 test sets of Speech-DF-Arena, outperforming all currently public systems on the leaderboard. It also obtains the lowest EER on five individual test sets and shows a favorable performance-complexity trade-off compared with larger leading systems. Overall, Teffic-Audio provides a strong and practical reference system for general speech deepfake detection.
Abstract (translated)
URL
https://arxiv.org/abs/2607.28351
PDF
-
Integrating Contextual Embeddings into Evaluation of Expressive MIDI Piano Performances
2026-07-30 09:22:54Dmitrii Gavrilev, Ilya Borovik, Vladimir ViroAbstract
Objective evaluation of expressive MIDI piano performances typically relies on attribute statistics such as timing, velocity, and duration of individual notes. However, these methods often disregard dependencies between notes, which poses a potential limitation in assessing the similarity between two sets of performances. In generative applications, the wide variety of expressive attributes makes it difficult to aggregate them into a single scalar metric for model selection. In this work, we reexamine attribute-scoped metrics and explore the perceptual properties of contextual embeddings from self-supervised symbolic music models, Aria and CLaMP3. Results from our listening study indicate that these models can be used as perceptual proxies, showing agreement with per-sample human ratings on par with traditional metrics. To measure conditional distributional similarity, we adapt Kernel Audio Distance to the symbolic music domain. Unlike Pearson correlation and reconstruction error, kernel-based methods on contextual embeddings do not require note alignment and are sensitive to contextual perturbations. To facilitate reproducibility, we release Pereval, an open-source library that integrates performance evaluation utilities, including both attribute-scoped and deep feature metrics.
Abstract (translated)
URL
https://arxiv.org/abs/2607.27909
PDF
-
CrowdioSet and PaRIRset: Two Datasets Towards Live Music Source Separation
2026-07-30 08:07:29Enric Gus\'o, Xavier SerraAbstract
Most Music Source Separation (MSS) models do not generalize well to live music recordings because they are trained on studio recordings alone, disregarding the venue acoustics, the speaker system's response and audience noise. We propose to bridge this gap by providing and training a model on two novel datasets. First, we present CrowdioSet: a noise dataset comprising 4800 real ambience tracks from Freesound and synthetic sing-alongs for the vocals in MUSDB18 and MOISESDB datasets, generated from zero-shot singing voice conversions. CrowdioSet enables effective audio denoising for live recordings, resulting in superior separation both in objective and subjective evaluations. Second, we introduce PaRIRset, a stereo impulse response dataset captured across 40 professional concert venues using a microphone array. Our results show that adding PaRIRset RIRs increases the performance of a MSS model compared to using real RIRs from Speech Enhancement tasks alone. We make the examples, code, model weights, PaRIRset, and CrowdioSet freely available to the public.
Abstract (translated)
URL
https://arxiv.org/abs/2607.27828
PDF
-
RIPPLE: Generating Multi-Channel Phase, Not Recovering It
Abstract
Generative models synthesize magnitude spectra with high fidelity, while phase is delegated to a recovery module---Griffin--Lim, a vocoder, or a latent decoder---applied independently to each channel. For multi-channel waveforms this delegation is costly: the physical content of spatial audio and three-component seismograms lives in the phase relationships between channels, precisely what channel-independent recovery cannot produce. The cost is also invisible, since the magnitude-based metrics common to both fields barely move when inter-channel phase coherence collapses---so a pipeline can discard the physical information in its output while still scoring well. We argue that phase should be generated, not recovered, and present RIPPLE (Rectified Inter-channel Phase with Prior-based LEarning), which reinterprets Griffin--Lim as a phase **prior** rather than a final estimator: initialized from the source phase, this prior carries the inter-channel structure to be preserved, and a rectified flow refines it toward the target under an explicit inter-channel phase loss. Tested on first-order ambisonics environment transfer and seismic cross-station translation---two physically unrelated domains---RIPPLE outperforms recovery-based pipelines on the coherence metrics that downstream analyses consume. The seismic case is decisive: across architecturally distinct generators, per-channel recovery leaves S-wave polarization error near the $57.3^\circ$ random expectation, whereas learned phase reduces it to $33.8^\circ$.
Abstract (translated)
URL
https://arxiv.org/abs/2607.27775
PDF
-
VocalRender: Score-Native Singing Voice Synthesis for Real-World Composition
2026-07-30 07:04:46Yukun Chen, Tianrui Wang, Zhaoxi Mu, Xinyu Yang, EngSiong ChngAbstract
Existing singing voice synthesis systems often require predefined durations, explicit duration prediction, or time-aligned acoustic guidance, which limits their compatibility with practical composition workflows. We propose VocalRender, a score-native system that directly synthesizes singing from lyrics, pitches, symbolic note values, and tempo. It uses an interleaved lyric--note representation and an autoregressive diffusion model to generate continuous acoustic latents while predicting the output length, eliminating the need for explicit duration prediction. Trained on a 2,300-hour singing dataset, VocalRender achieves strong intelligibility, strong melody control, and high speaker similarity across both in-domain and out-of-domain benchmarks. Notably, it outperforms the strongest baseline by $0.42$ points in naturalness CMOS, demonstrating the effectiveness of our proposed score-native architecture.
Abstract (translated)
URL
https://arxiv.org/abs/2607.27768
PDF
-
Improved Robustness in AI-Generated Music Detection
2026-07-29 20:43:13Emile Dugelay, Thomas Barand, Baptiste Campeas, Aur\'elien Laouar, Darius Afchar, Romain HennequinAbstract
AI music generators leave predictable spectral artifacts determined by their architecture. Existing detectors exploit these artifacts with near-perfect accuracy on raw generated tracks, but their performance collapses under simple audio manipulations, such as speed modification or pitch shifting. We address this open robustness problem by introducing a frequency-scaling-invariant detection pipeline that aims to prevent this kind of attack by design. Our method maps audio onto a log-frequency axis via a log-STFT remapping. A single learned cross-correlation filter, combined with max-pooling, provides shift invariance at inference time. Training uses a hybrid loss that jointly supervises binary detection and artifact-peak localization, regularizing boundary weights. Because robustness to speed change is built in by design, the detector is also interpretable: it outputs both a binary decision and an estimate of the applied speed-change factor.
Abstract (translated)
URL
https://arxiv.org/abs/2607.27454
PDF
-
MMAC: A Massive Multi-dimensional Benchmark for Audio Captioning
2026-07-29 16:38:08Weijie Wu, Junbo Li, Lin Li, Jun Fang, Qingyang HongAbstract
With the development of audio large language models (AudioLLMs), audio captioning needs to move from brief descriptions toward open-ended and fine-grained free-form descriptions. Existing evaluations often focus on generation quality or task performance, making it difficult to diagnose information coverage and description reliability. We propose MMAC, a \textbf{M}assive \textbf{M}ulti-dimensional benchmark for \textbf{A}udio \textbf{C}aptioning. MMAC contains 5,638 audio clips from more than 20 data sources, covering 6 capability categories and 15 evaluation dimensions. Given a model-generated caption, MMAC checks whether it mentions relevant information in the target dimension and whether the mentioned content is consistent with the reference label. We evaluate representative open-source and proprietary AudioLLMs. Results show clear differences across evaluation dimensions, information coverage, and description reliability. We will release the MMAC benchmark and evaluation code.
Abstract (translated)
URL
https://arxiv.org/abs/2607.27109
PDF
-
SKY-Piano: A Multimodal Piano Performance Dataset
2026-07-29 16:06:02Joonhyung Bae, Dawon Park, Taegyun Kwon, Yoon-Seok Choi, Hyeon Hur, Satoshi Obata, Shigeru Kai, Yohei Wada, Yu Takahashi, Akira Maezawa, Jaebum Park, Jonghwa Park, Juhan NamAbstract
Music information retrieval research on piano performance increasingly involves diverse modalities of data and annotations beyond audio and MIDI. We present SKY-Piano, a multimodal piano performance dataset that includes 11 hours of performance recordings of motion, multi-view video, audio, MIDI from 7 professional and 12 amateur pianists along with MusicXML scores. The performance pieces were selected considering playing technique, difficulty, and performer expertise on a shared core repertoire. The motion data include both hand and body motion, released in both flagged form, where samples lost to marker occlusion are marked as unreliable, and imputed form, where those gaps are reconstructed, together with Visual3D body-segment kinematics and other time-synchronized modalities. To easily browse different modalities of data at a glance, we provide an interactive web browser. In addition, we developed a fingering annotation model and tool for deriving pseudo fingering annotations from the MIDI and motion data. Lastly, we present MIDI-to-motion generation through a fine-tuning experiment as a use case of the dataset.
Abstract (translated)
URL
https://arxiv.org/abs/2607.27296
PDF
-
Detection of AI-generated stems within hybrid human-AI music
2026-07-29 13:03:58Fran\c{c}ois Rigaud, Gabriel Meseguer-Brocal, Benjamin Martin, Romain HennequinAbstract
This paper presents, to the best of our knowledge, the first study on detecting human-AI hybrid music tracks created by mixing human-produced and AI-generated stems. Building on recent work showing that AI music detectors can identify decoder-related artifacts in fully generated music, we investigate whether such artifacts remain detectable at the stem level after mixing. Using MUSDB18-HQ database in a two-stem vocals + accompaniment setting, we simulate hybrid mixtures by autoencoding individual stems with a neural codec. We compare two strategies combining AI-generated mix detection and source separation. A naive sequential pipeline, where source separation is followed by detection on separated sources, confirms that artifacts associated with an AI-generated stem are not reliably recovered by generic source separation systems. We therefore propose a parallel architecture in which source separation is only used to estimate source-relative energy within the mixture. We then train simple stem-specific binary classifiers that take as input the generated mix prediction together with the relative energy of the target stem on short audio chunks. Averaging chunk-level predictions yields encouraging track-level results, highlighting the potential of such approaches for detecting AI-generated stems in hybrid music.
Abstract (translated)
URL
https://arxiv.org/abs/2607.26874
PDF
-
Zero-Shot Face-to-Speech Synthesis via Latent Space Adaptation of a Style-Diffusion TTS Model
2026-07-29 10:33:56Carlos Mu\~noz-Romero, Jose A. Gonzalez-LopezAbstract
Zero-shot text-to-speech (TTS) clones a voice from a short audio prompt, but this reliance on reference audio is a barrier when only visual information is available, e.g. for historical figures or video-game characters. In this work, we propose a Face-to-Speech (F2S) framework that predicts a plausible voice from a static facial image. A lightweight Face Adapter, together with soft-tuning of the face encoder's upper blocks, aligns face-recognition features with the style space of a frozen StyleTTS 2 model, kept frozen during training. We evaluate on held-out identities from LRS3, a large-scale audiovisual corpus of English TED-talk videos. The synthesized speech is highly natural (UTMOS 3.7-4.0, matching or exceeding the 3.61 of ground truth), face-to-voice retrieval is consistently above chance, and the generated voice is consistent with the target speaker. Without any retraining, an English-trained adapter also produces fluent Spanish speech, indicating that the face-to-style mapping is largely language-agnostic.
Abstract (translated)
URL
https://arxiv.org/abs/2607.26742
PDF
-
Does EEG Foundation Models Transfer to Speech? A Benchmark on Overt and Imagined Speech Decoding
2026-07-29 10:06:01Owais Mujtaba Khanday, Mohamed Baha Ben Ticha, Sanae Belfrouh, Marc Ouellet, Jose A. Gonzalez-LopezAbstract
EEG foundation models pretrained on thousands of hours have shown large gains over task-specific networks for motor imagery, seizure detection, sleep staging, and emotion recognition, but their transfer to speech decoding-arguably the most demanding non-invasive BCI application-remains untested. We present the first systematic benchmark of EEG foundation models against strong convolutional baselines for speech decoding, using two corpora: UGR-MINDVOICE (overt and covert Iberian Spanish) and BCI Competition 2020 Track 3 (imagined speech). We compare two foundation models (LaBraM, EEGMamba) against three established baselines (EEGNet, ShallowFBCSPNet, EEGConformer) under a unified preprocessing and fine-tuning protocol. Large-scale EEG pretraining yields no consistent advantage over a 16K-parameter CNN on speech tasks, indicating that current general-purpose EEG pretraining does not yet transfer to speech production and motivating speech-specific foundation models.
Abstract (translated)
URL
https://arxiv.org/abs/2607.27268
PDF
-
MPEcho: A Melody and Phoneme-Aware Generative Framework for Controllable Cover Song Generation
2026-07-29 09:45:13Wei-Jaw Lee, Hsuan-Yu Yeh, Ting-Yi Hu, Chih-Pin Tan, Fang-Duo Tsai, Yi-Hsuan YangAbstract
Cover song generation (CSG) should preserve the melodic and linguistic content of a reference song while recreating the remaining musical components. The state-of-the-art model SongEcho utilizes $F_0$ sequences and voiced/unvoiced (V/UV) tags for conditioning; however, implicit linguistic information from V/UV tags cannot guarantee lyric accuracy, leading to a high phoneme error rate (PER). Inspired by singing voice synthesis (SVS), we propose MPEcho, which integrates a phoneme encoder and a length regulator (LR) into the SongEcho framework. By providing explicit phoneme-level conditioning and precise temporal boundaries, MPEcho significantly reduces PER. To enable this, we developed Phonsa, a Whisper-based automatic transcription model that provides high-precision phoneme-level annotations for singing voices, overcoming the scarcity of high-quality audio-phoneme pairs. Experimental results validate the effectiveness of Phonsa for alignment and MPEcho for end-to-end CSG. The audio samples, code and weights can be accessed from this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2607.26698
PDF
-
Few-Shot Open-Set Audio Classification via Transductive Prototype Refinement and Class Logit Enhancement
2026-07-29 08:29:47Tianyan Deng, Yanxiong Li, Rui Gao, Jiahao DuAbstract
Few-shot Open-set audio classification requires classifying query samples from known classes with a few labeled support samples while rejecting query samples from unknown classes. Transductive inference jointly observes the full unlabeled query set to improve prototype estimation, yet standard transductive updates do not distinguish known from unknown query samples, leaving prototypes vulnerable to open-set contamination. Drawing on latent-inlierness weighting and decoupled scoring for unknown-class samples, we propose a two-phase transductive method operating over a frozen audio encoder. First, each query sample is assigned a latent inlierness score that down-weights likely unknown-class samples, so that prototype refinement is driven primarily by known-class evidence. The refined prototypes are then directly optimized on a transductive loss combining support cross-entropy, inlierness-weighted conditional entropy minimization, and inlierness-weighted marginal entropy maximization, while open-set rejection uses a prior-adaptive free-energy score that adjusts its threshold with the prior proportion of unknown-class samples, decoupling detection from classification. Experiments on three audio datasets show our method achieves state-of-the-art results for few-shot open-set audio classification under multiple experimental conditions.
Abstract (translated)
URL
https://arxiv.org/abs/2607.26607
PDF
-
ThinkOmni: A Reasoning-Driven Omni-Modal LLM Framework for Audio Forgery Detection and Localization
2026-07-29 07:26:31Yuxiong Xu, Kaiqing Lin, Bin Li, Haodong Li, Sheng LiAbstract
Existing audio forgery detection and localization (AFDL) methods often overfit dataset-specific low-level artifacts, limiting their generalization to subtle, localized, and unseen manipulations. Recent audio large language model (ALLM)-based approaches cast AFDL as question answering but still model forensic evidence implicitly, without linking manipulation cues to predictions. To bridge this gap, we propose ThinkOmni, a reasoning-driven omni-modal large language model that jointly performs explicit forensic reasoning, spoofing detection, and temporal manipulation localization. To enable explicit reasoning supervision, we construct Forensic-Aware Chain-of-Thought (FACoT), a 100K-sample dataset with structured forensic evidence and reasoning annotations. Leveraging FACoT, we introduce Forensic-Aware Modality-Incremental Learning (FMIL), which progressively aligns semantic, acoustic, and spectral-visual representations with the LLM backbone to capture complementary forensic cues. We further propose Forensic-Consistent Multi-task Loss (FCML), which combines weighted cross-entropy with an adaptive localization loss to coordinate reasoning generation, spoofing detection, and temporal localization. Extensive experiments show that ThinkOmni achieves strong cross-dataset generalization in both detection and localization. Code, models, data, and inference examples are available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2607.26553
PDF
