CyberSec Research

Large-scale in-the-wild speech datasets have become more prevalent in recent years due to increased interest in models that can learn useful features from unlabelled data for tasks such as speech recognition or synthesis. These datasets often contain undesirable features, such as multiple speakers, non-target languages, and music, which may impact model learning. The Whilter model is proposed as a multitask solution to identify these undesirable samples. Whilter uses a Whisper encoder with an attention-based classifier to solve five diverse classification problems at once. In addition, an annotated dataset is published for a subset of two popular in-the-wild corpora. Whilter achieves F1 scores above 85% and equal error rates of 6.5% to 7.8% for three of five subtasks, outperforming a state-of-the-art BEATs classifier on speech-specific classes, with a notable decrease in processing time compared to a combination of single-task alternatives.

Steganalysis methods based on deep learning (DL) often struggle with computational complexity and challenges in generalizing across different datasets. Incorporating a graph neural network (GNN) into steganalysis schemes enables the leveraging of relational data for improved detection accuracy and adaptability. This paper presents the first application of a Graph Neural Network (GNN), specifically the GraphSAGE architecture, for steganalysis of compressed voice over IP (VoIP) speech streams. The method involves straightforward graph construction from VoIP streams and employs GraphSAGE to capture hierarchical steganalysis information, including both fine grained details and high level patterns, thereby achieving high detection accuracy. Experimental results demonstrate that the developed approach performs well in uncovering quantization index modulation (QIM)-based steganographic patterns in VoIP signals. It achieves detection accuracy exceeding 98 percent even for short 0.5 second samples, and 95.17 percent accuracy under challenging conditions with low embedding rates, representing an improvement of 2.8 percent over the best performing state of the art methods. Furthermore, the model exhibits superior efficiency, with an average detection time as low as 0.016 seconds for 0.5-second samples an improvement of 0.003 seconds. This makes it efficient for online steganalysis tasks, providing a superior balance between detection accuracy and efficiency under the constraint of short samples with low embedding rates.

End-to-end automatic speech recognition (ASR) systems based on transformer architectures, such as Whisper, offer high transcription accuracy and robustness. However, their autoregressive decoding is computationally expensive, hence limiting deployment on CPU-based and resource-constrained devices. Speculative decoding (SD) mitigates this issue by using a smaller draft model to propose candidate tokens, which are then verified by the main model. However, this approach is impractical for devices lacking hardware accelerators like GPUs. To address this, we propose \emph{Token Map Drafting}, a model-free SD technique that eliminates the need for a separate draft model. Instead, we leverage a precomputed n-gram token map derived from domain-specific training data, enabling efficient speculative decoding with minimal overhead. Our method significantly accelerates ASR inference in structured, low-perplexity domains without sacrificing transcription accuracy. Experimental results demonstrate decoding speed-ups of $1.27\times$ on the CI-AVSR dataset and $1.37\times$ on our internal dataset without degrading recognition accuracy. Additionally, our approach achieves a $10\%$ absolute improvement in decoding speed over the Distill-spec baseline running on CPU, highlighting its effectiveness for on-device ASR applications.

As speech generation technology advances, the risk of misuse through deepfake audio has become a pressing concern, which underscores the critical need for robust detection systems. However, many existing speech deepfake datasets are limited in scale and diversity, making it challenging to train models that can generalize well to unseen deepfakes. To address these gaps, we introduce SpeechFake, a large-scale dataset designed specifically for speech deepfake detection. SpeechFake includes over 3 million deepfake samples, totaling more than 3,000 hours of audio, generated using 40 different speech synthesis tools. The dataset encompasses a wide range of generation techniques, including text-to-speech, voice conversion, and neural vocoder, incorporating the latest cutting-edge methods. It also provides multilingual support, spanning 46 languages. In this paper, we offer a detailed overview of the dataset's creation, composition, and statistics. We also present baseline results by training detection models on SpeechFake, demonstrating strong performance on both its own test sets and various unseen test sets. Additionally, we conduct experiments to rigorously explore how generation methods, language diversity, and speaker variation affect detection performance. We believe SpeechFake will be a valuable resource for advancing speech deepfake detection and developing more robust models for evolving generation techniques.

Meaningful speech assessment is vital in clinical phonetics and therapy monitoring. This study examined the link between perceptual speech assessments and objective acoustic measures in a large head and neck cancer (HNC) dataset. Trained listeners provided ratings of intelligibility, articulation, voice quality, phonation, speech rate, nasality, and background noise on speech. Strong correlations were found between subjective intelligibility, articulation, and voice quality, likely due to a shared underlying cause of speech symptoms in our speaker population. Objective measures of intelligibility and speech rate aligned with their subjective counterpart. Our results suggest that a single intelligibility measure may be sufficient for the clinical monitoring of speakers treated for HNC using concomitant chemoradiation.

Multimodal emotion recognition (MER) is crucial for enabling emotionally intelligent systems that perceive and respond to human emotions. However, existing methods suffer from limited cross-modal interaction and imbalanced contributions across modalities. To address these issues, we propose Sync-TVA, an end-to-end graph-attention framework featuring modality-specific dynamic enhancement and structured cross-modal fusion. Our design incorporates a dynamic enhancement module for each modality and constructs heterogeneous cross-modal graphs to model semantic relations across text, audio, and visual features. A cross-attention fusion mechanism further aligns multimodal cues for robust emotion inference. Experiments on MELD and IEMOCAP demonstrate consistent improvements over state-of-the-art models in both accuracy and weighted F1 score, especially under class-imbalanced conditions.

This study presented the development of a deep learning-based Automatic Speech Recognition system for Shona, a low-resource language characterized by unique tonal and grammatical complexities. The research aimed to address the challenges posed by limited training data, lack of labelled data, and the intricate tonal nuances present in Shona speech, with the objective of achieving significant improvements in recognition accuracy compared to traditional statistical models. The research first explored the feasibility of using deep learning to develop an accurate ASR system for Shona. Second, it investigated the specific challenges involved in designing and implementing deep learning architectures for Shona speech recognition and proposed strategies to mitigate these challenges. Lastly, it compared the performance of the deep learning-based model with existing statistical models in terms of accuracy. The developed ASR system utilized a hybrid architecture consisting of a Convolutional Neural Network for acoustic modelling and a Long Short-Term Memory network for language modelling. To overcome the scarcity of data, data augmentation techniques and transfer learning were employed. Attention mechanisms were also incorporated to accommodate the tonal nature of Shona speech. The resulting ASR system achieved impressive results, with a Word Error Rate of 29%, Phoneme Error Rate of 12%, and an overall accuracy of 74%. These metrics indicated the potential of deep learning to enhance ASR accuracy for under-resourced languages like Shona. This study contributed to the advancement of ASR technology for under-resourced languages like Shona, ultimately fostering improved accessibility and communication for Shona speakers worldwide.

We present Music Arena, an open platform for scalable human preference evaluation of text-to-music (TTM) models. Soliciting human preferences via listening studies is the gold standard for evaluation in TTM, but these studies are expensive to conduct and difficult to compare, as study protocols may differ across systems. Moreover, human preferences might help researchers align their TTM systems or improve automatic evaluation metrics, but an open and renewable source of preferences does not currently exist. We aim to fill these gaps by offering *live* evaluation for TTM. In Music Arena, real-world users input text prompts of their choosing and compare outputs from two TTM systems, and their preferences are used to compile a leaderboard. While Music Arena follows recent evaluation trends in other AI domains, we also design it with key features tailored to music: an LLM-based routing system to navigate the heterogeneous type signatures of TTM systems, and the collection of *detailed* preferences including listening data and natural language feedback. We also propose a rolling data release policy with user privacy guarantees, providing a renewable source of preference data and increasing platform transparency. Through its standardized evaluation protocol, transparent data access policies, and music-specific features, Music Arena not only addresses key challenges in the TTM ecosystem but also demonstrates how live evaluation can be thoughtfully adapted to unique characteristics of specific AI domains. Music Arena is available at: https://music-arena.org

Diffusion and flow-matching models have revolutionized automatic text-to-audio generation in recent times. These models are increasingly capable of generating high quality and faithful audio outputs capturing to speech and acoustic events. However, there is still much room for improvement in creative audio generation that primarily involves music and songs. Recent open lyrics-to-song models, such as, DiffRhythm, ACE-Step, and LeVo, have set an acceptable standard in automatic song generation for recreational use. However, these models lack fine-grained word-level controllability often desired by musicians in their workflows. To the best of our knowledge, our flow-matching-based JAM is the first effort toward endowing word-level timing and duration control in song generation, allowing fine-grained vocal control. To enhance the quality of generated songs to better align with human preferences, we implement aesthetic alignment through Direct Preference Optimization, which iteratively refines the model using a synthetic dataset, eliminating the need or manual data annotations. Furthermore, we aim to standardize the evaluation of such lyrics-to-song models through our public evaluation dataset JAME. We show that JAM outperforms the existing models in terms of the music-specific attributes.

Selecting appropriate inductive biases is an essential step in the design of machine learning models, especially when working with audio, where even short clips may contain millions of samples. To this end, we propose the combolutional layer: a learned-delay IIR comb filter and fused envelope detector, which extracts harmonic features in the time domain. We demonstrate the efficacy of the combolutional layer on three information retrieval tasks, evaluate its computational cost relative to other audio frontends, and provide efficient implementations for training. We find that the combolutional layer is an effective replacement for convolutional layers in audio tasks where precise harmonic analysis is important, e.g., piano transcription, speaker classification, and key detection. Additionally, the combolutional layer has several other key benefits over existing frontends, namely: low parameter count, efficient CPU inference, strictly real-valued computations, and improved interpretability.

Despite the rapid development of neural vocoders in recent years, they usually suffer from some intrinsic challenges like opaque modeling, and parameter-performance trade-off. In this study, we propose an innovative time-frequency (T-F) domain-based neural vocoder to resolve the above-mentioned challenges. To be specific, we bridge the connection between the classical signal range-null decomposition (RND) theory and vocoder task, and the reconstruction of target spectrogram can be decomposed into the superimposition between the range-space and null-space, where the former is enabled by a linear domain shift from the original mel-scale domain to the target linear-scale domain, and the latter is instantiated via a learnable network for further spectral detail generation. Accordingly, we propose a novel dual-path framework, where the spectrum is hierarchically encoded/decoded, and the cross- and narrow-band modules are elaborately devised for efficient sub-band and sequential modeling. Comprehensive experiments are conducted on the LJSpeech and LibriTTS benchmarks. Quantitative and qualitative results show that while enjoying lightweight network parameters, the proposed approach yields state-of-the-art performance among existing advanced methods. Our code and the pretrained model weights are available at https://github.com/Andong-Li-speech/RNDVoC.

This paper proposes a method for generating machine-type-specific anomalies to evaluate the relative performance of unsupervised anomalous sound detection (UASD) systems across different machine types, even in the absence of real anomaly sound data. Conventional keyword-based data augmentation methods often produce unrealistic sounds due to their reliance on manually defined labels, limiting scalability as machine types and anomaly patterns diversify. Advanced audio generative models, such as MIMII-Gen, show promise but typically depend on anomalous training data, making them less effective when diverse anomalous examples are unavailable. To address these limitations, we propose a novel synthesis approach leveraging large language models (LLMs) to interpret textual descriptions of faults and automatically select audio transformation functions, converting normal machine sounds into diverse and plausible anomalous sounds. We validate this approach by evaluating a UASD system trained only on normal sounds from five machine types, using both real and synthetic anomaly data. Experimental results reveal consistent trends in relative detection difficulty across machine types between synthetic and real anomalies. This finding supports our hypothesis and highlights the effectiveness of the proposed LLM-based synthesis approach for relative evaluation of UASD systems.

Music enhances video narratives and emotions, driving demand for automatic video-to-music (V2M) generation. However, existing V2M methods relying solely on visual features or supplementary textual inputs generate music in a black-box manner, often failing to meet user expectations. To address this challenge, we propose a novel multi-condition guided V2M generation framework that incorporates multiple time-varying conditions for enhanced control over music generation. Our method uses a two-stage training strategy that enables learning of V2M fundamentals and audiovisual temporal synchronization while meeting users' needs for multi-condition control. In the first stage, we introduce a fine-grained feature selection module and a progressive temporal alignment attention mechanism to ensure flexible feature alignment. For the second stage, we develop a dynamic conditional fusion module and a control-guided decoder module to integrate multiple conditions and accurately guide the music composition process. Extensive experiments demonstrate that our method outperforms existing V2M pipelines in both subjective and objective evaluations, significantly enhancing control and alignment with user expectations.

Audio effects (AFXs) are essential tools in music production, frequently applied in chains to shape timbre and dynamics. The order of AFXs in a chain plays a crucial role in determining the final sound, particularly when non-linear (e.g., distortion) or time-variant (e.g., chorus) processors are involved. Despite its importance, most AFX-related studies have primarily focused on estimating effect types and their parameters from a wet signal. To address this gap, we formulate AFX chain recognition as the task of jointly estimating AFX types and their order from a wet signal. We propose a neural-network-based method that embeds wet signals into a hyperbolic space and classifies their AFX chains. Hyperbolic space can represent tree-structured data more efficiently than Euclidean space due to its exponential expansion property. Since AFX chains can be represented as trees, with AFXs as nodes and edges encoding effect order, hyperbolic space is well-suited for modeling the exponentially growing and non-commutative nature of ordered AFX combinations, where changes in effect order can result in different final sounds. Experiments using guitar sounds demonstrate that, with an appropriate curvature, the proposed method outperforms its Euclidean counterpart. Further analysis based on AFX type and chain length highlights the effectiveness of the proposed method in capturing AFX order.

This paper introduces Binaural Sound Event Localization and Detection (BiSELD), a task that aims to jointly detect and localize multiple sound events using binaural audio, inspired by the spatial hearing mechanism of humans. To support this task, we present a synthetic benchmark dataset, called the Binaural Set, which simulates realistic auditory scenes using measured head-related transfer functions (HRTFs) and diverse sound events. To effectively address the BiSELD task, we propose a new input feature representation called the Binaural Time-Frequency Feature (BTFF), which encodes interaural time difference (ITD), interaural level difference (ILD), and high-frequency spectral cues (SC) from binaural signals. BTFF is composed of eight channels, including left and right mel-spectrograms, velocity-maps, SC-maps, and ITD-/ILD-maps, designed to cover different spatial cues across frequency bands and spatial axes. A CRNN-based model, BiSELDnet, is then developed to learn both spectro-temporal patterns and HRTF-based localization cues from BTFF. Experiments on the Binaural Set show that each BTFF sub-feature enhances task performance: V-map improves detection, ITD-/ILD-maps enable accurate horizontal localization, and SC-map captures vertical spatial cues. The final system achieves a SELD error of 0.110 with 87.1% F-score and 4.4{\deg} localization error, demonstrating the effectiveness of the proposed framework in mimicking human-like auditory perception.

Recent advances in synthetic speech have made audio deepfakes increasingly realistic, posing significant security risks. Existing detection methods that rely on a single modality, either raw waveform embeddings or spectral based features, are vulnerable to non spoof disturbances and often overfit to known forgery algorithms, resulting in poor generalization to unseen attacks. To address these shortcomings, we investigate hybrid fusion frameworks that integrate self supervised learning (SSL) based representations with handcrafted spectral descriptors (MFCC , LFCC, CQCC). By aligning and combining complementary information across modalities, these fusion approaches capture subtle artifacts that single feature approaches typically overlook. We explore several fusion strategies, including simple concatenation, cross attention, mutual cross attention, and a learnable gating mechanism, to optimally blend SSL features with fine grained spectral cues. We evaluate our approach on four challenging public benchmarks and report generalization performance. All fusion variants consistently outperform an SSL only baseline, with the cross attention strategy achieving the best generalization with a 38% relative reduction in equal error rate (EER). These results confirm that joint modeling of waveform and spectral views produces robust, domain agnostic representations for audio deepfake detection.

Recent audio LLMs have emerged rapidly, demonstrating strong generalization across various speech tasks. However, given the inherent complexity of speech signals, these models inevitably suffer from performance degradation in specific target domains. To address this, we focus on enhancing audio LLMs in target domains without any labeled data. We propose a self-improvement method called SI-SDA, leveraging the information embedded in large-model decoding to evaluate the quality of generated pseudo labels and then perform domain adaptation based on reinforcement learning optimization. Experimental results show that our method consistently and significantly improves audio LLM performance, outperforming existing baselines in WER and BLEU across multiple public datasets of automatic speech recognition (ASR), spoken question-answering (SQA), and speech-to-text translation (S2TT). Furthermore, our approach exhibits high data efficiency, underscoring its potential for real-world deployment.

The rapid advancement of Zero-Shot Text-to-Speech (ZS-TTS) technology has enabled high-fidelity voice synthesis from minimal audio cues, raising significant privacy and ethical concerns. Despite the threats to voice privacy, research to selectively remove the knowledge to replicate unwanted individual voices from pre-trained model parameters has not been explored. In this paper, we address the new challenge of speaker identity unlearning for ZS-TTS systems. To meet this goal, we propose the first machine unlearning frameworks for ZS-TTS, especially Teacher-Guided Unlearning (TGU), designed to ensure the model forgets designated speaker identities while retaining its ability to generate accurate speech for other speakers. Our proposed methods incorporate randomness to prevent consistent replication of forget speakers' voices, assuring unlearned identities remain untraceable. Additionally, we propose a new evaluation metric, speaker-Zero Retrain Forgetting (spk-ZRF). This assesses the model's ability to disregard prompts associated with forgotten speakers, effectively neutralizing its knowledge of these voices. The experiments conducted on the state-of-the-art model demonstrate that TGU prevents the model from replicating forget speakers' voices while maintaining high quality for other speakers. The demo is available at https://speechunlearn.github.io/