Bridging the SEA Gap: An Initial Benchmark for Neural Audio Codec-Synthesized Speech Deepfakes in South-East Asian Languages

Codecfakes (CFs) are a type of speech deepfakes generated through Audio Language Models (ALMs), with Neural Audio Codecs (NACs) forming the core mechanism for speech encoding and generation. CFs exhibit distributional characteristics that differ from vocoder-based deepfakes, causing detectors trained on vocoder data to generalize poorly to CFs detection. Although this has led to the development of CF detection benchmarks, existing resources are largely confined to English -- and to a limited extent Chinese -- leaving South-East Asian (SEA) languages unexplored. To bridge this gap, we introduce SEA-CF, the first large-scale benchmark for CF detection spanning multiple SEA languages, diverse speaker profiles, and a wide range of NAC architectures. SEA-CF is constructed by synthesizing publicly available real speech corpora. Our experiments show that state-of-the-art (SOTA) CF detectors trained on English-centric datasets fail to generalize to SEA speech due to language-specific phonetic structures, tonal variations, and rich prosodic diversity. We further conduct a comprehensive zero-shot and fine-tuned evaluation of recent SOTA ALMs on SEA-CF. Fine-tuning the ALMs improves performance, however, these are very large being impractical for real-world application due to their scale, particularly in low-resource and latency-constrained settings. To address this limitation, we propose a novel small-ALM, GARUDA tailored for CF detection, which delivers strong performance while remaining lightweight. Extensive evaluations demonstrate that the proposed Small-ALM outperforms strong end-to-end and ALM-based baselines, establishing a new, practical direction for robust CF detection in SEA languages and beyond.

Indic-CodecFake meets SATYAM: Towards Detecting Neural Audio Codec Synthesized Speech Deepfakes in Indic Languages

The rapid advancement of Audio Large Language Models (ALMs), driven by Neural Audio Codecs (NACs), has led to the emergence of highly realistic speech deepfakes, commonly referred to as CodecFakes (CFs). Consequently, CF detection has attracted increasing attention from the research community. However, existing studies predominantly focus on English or Chinese, leaving the vulnerability of Indic languages largely unexplored. To bridge this gap, we introduce Indic-CodecFake (ICF) dataset, the first large-scale benchmark comprising real and NAC-synthesized speech across multiple Indic languages, diverse speaker profiles, and multiple NAC types. We use IndicSUPERB as the real speech corpus for generation of ICF dataset. Our experiments demonstrate that state-of-the-art (SOTA) CF detectors trained on English-centric datasets fail to generalize to ICF, underscoring the challenges posed by phonetic diversity and prosodic variability in Indic speech. Further, we present systematic evaluation of SOTA ALMs in a zero-shot setting on ICF dataset. We evaluate these ALMs as they have shown effectiveness for different speech tasks. However, our findings reveal that current ALMs exhibit consistently poor performance. To address this, we propose SATYAM, a novel hyperbolic ALM tailored for CF detection in Indic languages. SATYAM integrates semantic representations from Whisper and prosodic representations from TRILLsson using through Bhattacharya distance in hyperbolic space and subsequently performs the same alignment procedure between the fused speech representation and an input conditioning prompt. This dual-stage fusion framework enables SATYAM to effectively model hierarchical relationships both within speech (semantic-prosodic) and across modalities (speech-text). Extensive evaluations show that SATYAM consistently outperforms competitive end-to-end and ALM-based baselines on the ICF benchmark.

Are Multimodal Foundation Models All That Is Needed for Emofake Detection?

In this work, we investigate multimodal foundation models (MFMs) for EmoFake detection (EFD) and hypothesize that they will outperform audio foundation models (AFMs). MFMs due to their cross-modal pre-training, learns emotional patterns from multiple modalities, while AFMs rely only on audio. As such, MFMs can better recognize unnatural emotional shifts and inconsistencies in manipulated audio, making them more effective at distinguishing real from fake emotional expressions. To validate our hypothesis, we conduct a comprehensive comparative analysis of state-of-the-art (SOTA) MFMs (e.g. LanguageBind) alongside AFMs (e.g. WavLM). Our experiments confirm that MFMs surpass AFMs for EFD. Beyond individual foundation models (FMs) performance, we explore FMs fusion, motivated by findings in related research areas such synthetic speech detection and speech emotion recognition. To this end, we propose SCAR, a novel framework for effective fusion. SCAR introduces a nested cross-attention mechanism, where representations from FMs interact at two stages sequentially to refine information exchange. Additionally, a self-attention refinement module further enhances feature representations by reinforcing important cross-FM cues while suppressing noise. Through SCAR with synergistic fusion of MFMs, we achieve SOTA performance, surpassing both standalone FMs and conventional fusion approaches and previous works on EFD.

Rethinking Cross-Corpus Speech Emotion Recognition Benchmarking: Are Paralinguistic Pre-Trained Representations Sufficient?

Recent benchmarks evaluating pre-trained models (PTMs) for cross-corpus speech emotion recognition (SER) have overlooked PTM pre-trained for paralinguistic speech processing (PSP), raising concerns about their reliability, since SER is inherently a paralinguistic task. We hypothesize that PSP-focused PTM will perform better in cross-corpus SER settings. To test this, we analyze state-of-the-art PTMs representations including paralinguistic, monolingual, multilingual, and speaker recognition. Our results confirm that TRILLsson (a paralinguistic PTM) outperforms others, reinforcing the need to consider PSP-focused PTMs in cross-corpus SER benchmarks. This study enhances benchmark trustworthiness and guides PTMs evaluations for reliable cross-corpus SER.

Towards Neural Audio Codec Source Parsing

A new class of audio deepfakes-codecfakes (CFs)-has recently caught attention, synthesized by Audio Language Models that leverage neural audio codecs (NACs) in the backend. In response, the community has introduced dedicated benchmarks and tailored detection strategies. As the field advances, efforts have moved beyond binary detection toward source attribution, including open-set attribution, which aims to identify the NAC responsible for generation and flag novel, unseen ones during inference. This shift toward source attribution improves forensic interpretability and accountability. However, open-set attribution remains fundamentally limited: while it can detect that a NAC is unfamiliar, it cannot characterize or identify individual unseen codecs. It treats such inputs as generic ``unknowns'', lacking insight into their internal configuration. This leads to major shortcomings: limited generalization to new NACs and inability to resolve fine-grained variations within NAC families. To address these gaps, we propose Neural Audio Codec Source Parsing (NACSP) - a paradigm shift that reframes source attribution for CFs as structured regression over generative NAC parameters such as quantizers, bandwidth, and sampling rate. We formulate NACSP as a multi-task regression task for predicting these NAC parameters and establish the first comprehensive benchmark using various state-of-the-art speech pre-trained models (PTMs). To this end, we propose HYDRA, a novel framework that leverages hyperbolic geometry to disentangle complex latent properties from PTM representations. By employing task-specific attention over multiple curvature-aware hyperbolic subspaces, HYDRA enables superior multi-task generalization. Our extensive experiments show HYDRA achieves top results on benchmark CFs datasets compared to baselines operating in Euclidean space.

Investigating Prosodic Signatures via Speech Pre-Trained Models for Audio Deepfake Source Attribution

In this work, we investigate various state-of-the-art (SOTA) speech pre-trained models (PTMs) for their capability to capture prosodic signatures of the generative sources for audio deepfake source attribution (ADSD). These prosodic characteristics can be considered one of major signatures for ADSD, which is unique to each source. So better is the PTM at capturing prosodic signs better the ADSD performance. We consider various SOTA PTMs that have shown top performance in different prosodic tasks for our experiments on benchmark datasets, ASVSpoof 2019 and CFAD. x-vector (speaker recognition PTM) attains the highest performance in comparison to all the PTMs considered despite consisting lowest model parameters. This higher performance can be due to its speaker recognition pre-training that enables it for capturing unique prosodic characteristics of the sources in a better way. Further, motivated from tasks such as audio deepfake detection and speech recognition, where fusion of PTMs representations lead to improved performance, we explore the same and propose FINDER for effective fusion of such representations. With fusion of Whisper and x-vector representations through FINDER, we achieved the topmost performance in comparison to all the individual PTMs as well as baseline fusion techniques and attaining SOTA performance.

Beyond Speech and More: Investigating the Emergent Ability of Speech Foundation Models for Classifying Physiological Time-Series Signals

Despite being trained exclusively on speech data, speech foundation models (SFMs) like Whisper have shown impressive performance in non-speech tasks such as audio classification. This is partly because speech shares some common traits with audio, enabling SFMs to transfer effectively. In this study, we push the boundaries by evaluating SFMs on a more challenging out-of-domain (OOD) task: classifying physiological time-series signals. We test two key hypotheses: first, that SFMs can generalize to physiological signals by capturing shared temporal patterns; second, that multilingual SFMs will outperform others due to their exposure to greater variability during pre-training, leading to more robust, generalized representations. Our experiments, conducted for stress recognition using ECG (Electrocardiogram), EMG (Electromyography), and EDA (Electrodermal Activity) signals, reveal that models trained on SFM-derived representations outperform those trained on raw physiological signals. Among all models, multilingual SFMs achieve the highest accuracy, supporting our hypothesis and demonstrating their OOD capabilities. This work positions SFMs as promising tools for new uncharted domains beyond speech.

SeQuiFi: Mitigating Catastrophic Forgetting in Speech Emotion Recognition with Sequential Class-Finetuning

In this work, we introduce SeQuiFi, a novel approach for mitigating catastrophic forgetting (CF) in speech emotion recognition (SER). SeQuiFi adopts a sequential class-finetuning strategy, where the model is fine-tuned incrementally on one emotion class at a time, preserving and enhancing retention for each class. While various state-of-the-art (SOTA) methods, such as regularization-based, memory-based, and weight-averaging techniques, have been proposed to address CF, it still remains a challenge, particularly with diverse and multilingual datasets. Through extensive experiments, we demonstrate that SeQuiFi significantly outperforms both vanilla fine-tuning and SOTA continual learning techniques in terms of accuracy and F1 scores on multiple benchmark SER datasets, including CREMA-D, RAVDESS, Emo-DB, MESD, and SHEMO, covering different languages.

Multi-View Multi-Task Modeling with Speech Foundation Models for Speech Forensic Tasks

Speech forensic tasks (SFTs), such as automatic speaker recognition (ASR), speech emotion recognition (SER), gender recognition (GR), and age estimation (AE), find use in different security and biometric applications. Previous works have applied various techniques, with recent studies focusing on applying speech foundation models (SFMs) for improved performance. However, most prior efforts have centered on building individual models for each task separately, despite the inherent similarities among these tasks. This isolated approach results in higher computational resource requirements, increased costs, time consumption, and maintenance challenges. In this study, we address these challenges by employing a multi-task learning strategy. Firstly, we explore the various state-of-the-art (SOTA) SFMs by extracting their representations for learning these SFTs and investigating their effectiveness at each task specifically. Secondly, we analyze the performance of the extracted representations on the SFTs in a multi-task learning framework. We observe a decline in performance when SFTs are modeled together compared to individual task-specific models, and as a remedy, we propose multi-view learning (MVL). Views are representations from different SFMs transformed into distinct abstract spaces by characteristics unique to each SFM. By leveraging MVL, we integrate these diverse representations to capture complementary information across tasks, enhancing the shared learning process. We introduce a new framework called TANGO (Task Alignment with iNter-view Gated Optimal transport) to implement this approach. With TANGO, we achieve the topmost performance in comparison to individual SFM representations as well as baseline fusion techniques across benchmark datasets such as CREMA-D, emo-DB, and BAVED.

Representation Loss Minimization with Randomized Selection Strategy for Efficient Environmental Fake Audio Detection

The adaptation of foundation models has significantly advanced environmental audio deepfake detection (EADD), a rapidly growing area of research. These models are typically fine-tuned or utilized in their frozen states for downstream tasks. However, the dimensionality of their representations can substantially lead to a high parameter count of downstream models, leading to higher computational demands. So, a general way is to compress these representations by leveraging state-of-the-art (SOTA) unsupervised dimensionality reduction techniques (PCA, SVD, KPCA, GRP) for efficient EADD. However, with the application of such techniques, we observe a drop in performance. So in this paper, we show that representation vectors contain redundant information, and randomly selecting 40-50% of representation values and building downstream models on it preserves or sometimes even improves performance. We show that such random selection preserves more performance than the SOTA dimensionality reduction techniques while reducing model parameters and inference time by almost over half.