CyberSec Research

We present a geometry-driven method for normalizing dysarthric speech using local Lie group transformations of spectrograms. Time, frequency, and amplitude distortions are modeled as smooth, invertible deformations, parameterized by scalar fields and applied via exponential maps. A neural network is trained to infer these fields from synthetic distortions of typical speech-without using any pathological data. At test time, the model applies an approximate inverse to real dysarthric inputs. Despite zero-shot generalization, we observe substantial ASR gains, including up to 16 percentage points WER reduction on challenging TORGO samples, with no degradation on clean speech. This work introduces a principled, interpretable approach for robust speech recognition under motor speech disorders

Voice conversion is a task of synthesizing an utterance with target speaker's voice while maintaining linguistic information of the source utterance. While a speaker can produce varying utterances from a single script with different intonations, conventional voice conversion models were limited to producing only one result per source input. To overcome this limitation, we propose a novel approach for voice conversion with diverse intonations using conditional variational autoencoder (CVAE). Experiments have shown that the speaker's style feature can be mapped into a latent space with Gaussian distribution. We have also been able to convert voices with more diverse intonation by making the posterior of the latent space more complex with inverse autoregressive flow (IAF). As a result, the converted voice not only has a diversity of intonations, but also has better sound quality than the model without CVAE.

The large integration of microphones into devices increases the opportunities for Acoustic Side-Channel Attacks (ASCAs), as these can be used to capture keystrokes' audio signals that might reveal sensitive information. However, the current State-Of-The-Art (SOTA) models for ASCAs, including Convolutional Neural Networks (CNNs) and hybrid models, such as CoAtNet, still exhibit limited robustness under realistic noisy conditions. Solving this problem requires either: (i) an increased model's capacity to infer contextual information from longer sequences, allowing the model to learn that an initially noisily typed word is the same as a futurely collected non-noisy word, or (ii) an approach to fix misidentified information from the contexts, as one does not type random words, but the ones that best fit the conversation context. In this paper, we demonstrate that both strategies are viable and complementary solutions for making ASCAs practical. We observed that no existing solution leverages advanced transformer architectures' power for these tasks and propose that: (i) Visual Transformers (VTs) are the candidate solutions for capturing long-term contextual information and (ii) transformer-powered Large Language Models (LLMs) are the candidate solutions to fix the ``typos'' (mispredictions) the model might make. Thus, we here present the first-of-its-kind approach that integrates VTs and LLMs for ASCAs. We first show that VTs achieve SOTA performance in classifying keystrokes when compared to the previous CNN benchmark. Second, we demonstrate that LLMs can mitigate the impact of real-world noise. Evaluations on the natural sentences revealed that: (i) incorporating LLMs (e.g., GPT-4o) in our ASCA pipeline boosts the performance of error-correction tasks; and (ii) the comparable performance can be attained by a lightweight, fine-tuned smaller LLM (67 times smaller than GPT-4o), using...

Audiobook generation, which creates vivid and emotion-rich audio works, faces challenges in conveying complex emotions, achieving human-like qualities, and aligning evaluations with human preferences. Existing text-to-speech (TTS) methods are often limited to specific scenarios, struggle with emotional transitions, and lack automatic human-aligned evaluation benchmarks, instead relying on either misaligned automated metrics or costly human assessments. To address these issues, we propose Dopamine Audiobook, a new unified training-free system leveraging a multimodal large language model (MLLM) as an AI agent for emotional and human-like audiobook generation and evaluation. Specifically, we first design a flow-based emotion-enhanced framework that decomposes complex emotional speech synthesis into controllable sub-tasks. Then, we propose an adaptive model selection module that dynamically selects the most suitable TTS methods from a set of existing state-of-the-art (SOTA) TTS methods for diverse scenarios. We further enhance emotional expressiveness through paralinguistic augmentation and prosody retrieval at word and utterance levels. For evaluation, we propose a novel GPT-based evaluation framework incorporating self-critique, perspective-taking, and psychological MagicEmo prompts to ensure human-aligned and self-aligned assessments. Experiments show that our method generates long speech with superior emotional expression to SOTA TTS models in various metrics. Importantly, our evaluation framework demonstrates better alignment with human preferences and transferability across audio tasks. Project website with audio samples can be found at https://dopamine-audiobook.github.io.

Rich-text captions are essential to help communication for Deaf and hard-of-hearing (DHH) people, second-language learners, and those with autism spectrum disorder (ASD). They also preserve nuances when converting speech to text, enhancing the realism of presentation scripts and conversation or speech logs. However, current real-time captioning systems lack the capability to alter text attributes (ex. capitalization, sizes, and fonts) at the word level, hindering the accurate conveyance of speaker intent that is expressed in the tones or intonations of the speech. For example, ''YOU should do this'' tends to be considered as indicating ''You'' as the focus of the sentence, whereas ''You should do THIS'' tends to be ''This'' as the focus. This paper proposes a solution that changes the text decorations at the word level in real time. As a prototype, we developed an application that adjusts word size based on the loudness of each spoken word. Feedback from users implies that this system helped to convey the speaker's intent, offering a more engaging and accessible captioning experience.

Music editing is an important step in music production, which has broad applications, including game development and film production. Most existing zero-shot text-guided methods rely on pretrained diffusion models by involving forward-backward diffusion processes for editing. However, these methods often struggle to maintain the music content consistency. Additionally, text instructions alone usually fail to accurately describe the desired music. In this paper, we propose two music editing methods that enhance the consistency between the original and edited music by leveraging score distillation. The first method, SteerMusic, is a coarse-grained zero-shot editing approach using delta denoising score. The second method, SteerMusic+, enables fine-grained personalized music editing by manipulating a concept token that represents a user-defined musical style. SteerMusic+ allows for the editing of music into any user-defined musical styles that cannot be achieved by the text instructions alone. Experimental results show that our methods outperform existing approaches in preserving both music content consistency and editing fidelity. User studies further validate that our methods achieve superior music editing quality. Audio examples are available on https://steermusic.pages.dev/.

This study investigates the classification of progressive rock music, a genre characterized by complex compositions and diverse instrumentation, distinct from other musical styles. Addressing this Music Information Retrieval (MIR) task, we extracted comprehensive audio features, including spectrograms, Mel-Frequency Cepstral Coefficients (MFCCs), chromagrams, and beat positions from song snippets using the Librosa library. A winner-take-all voting strategy was employed to aggregate snippet-level predictions into final song classifications. We conducted a comparative analysis of various machine learning techniques. Ensemble methods, encompassing Bagging (Random Forest, ExtraTrees, Bagging Classifier) and Boosting (XGBoost, Gradient Boosting), were explored, utilizing Principal Component Analysis (PCA) for dimensionality reduction to manage computational constraints with high-dimensional feature sets. Additionally, deep learning approaches were investigated, including the development of custom 1D Convolutional Neural Network (1D CNN) architectures (named "Zuck" and "Satya") featuring specific layer configurations, normalization, and activation functions. Furthermore, we fine-tuned a state-of-the-art Audio Spectrogram Transformer (AST) model, leveraging its attention-based mechanisms for audio classification. Performance evaluation on validation and test sets revealed varying effectiveness across models, with ensemble methods like Extra Trees achieving test accuracies up to 76.38%. This research provides insights into the application and relative performance of diverse machine learning paradigms for the nuanced task of progressive rock genre classification.

Generalizability, the capacity of a robust model to perform effectively on unseen data, is crucial for audio deepfake detection due to the rapid evolution of text-to-speech (TTS) and voice conversion (VC) technologies. A promising approach to differentiate between bonafide and spoof samples lies in identifying intrinsic disparities to enhance model generalizability. From an information-theoretic perspective, we hypothesize the information content is one of the intrinsic differences: bonafide sample represents a dense, information-rich sampling of the real world, whereas spoof sample is typically derived from lower-dimensional, less informative representations. To implement this, we introduce frame-level latent information entropy detector(f-InfoED), a framework that extracts distinctive information entropy from latent representations at the frame level to identify audio deepfakes. Furthermore, we present AdaLAM, which extends large pre-trained audio models with trainable adapters for enhanced feature extraction. To facilitate comprehensive evaluation, the audio deepfake forensics 2024 (ADFF 2024) dataset was built by the latest TTS and VC methods. Extensive experiments demonstrate that our proposed approach achieves state-of-the-art performance and exhibits remarkable generalization capabilities. Further analytical studies confirms the efficacy of AdaLAM in extracting discriminative audio features and f-InfoED in leveraging latent entropy information for more generalized deepfake detection.

Imagine placing your smartphone on a table in a noisy restaurant and clearly capturing the voices of friends seated around you, or recording a lecturer's voice with clarity in a reverberant auditorium. We introduce SonicSieve, the first intelligent directional speech extraction system for smartphones using a bio-inspired acoustic microstructure. Our passive design embeds directional cues onto incoming speech without any additional electronics. It attaches to the in-line mic of low-cost wired earphones which can be attached to smartphones. We present an end-to-end neural network that processes the raw audio mixtures in real-time on mobile devices. Our results show that SonicSieve achieves a signal quality improvement of 5.0 dB when focusing on a 30{\deg} angular region. Additionally, the performance of our system based on only two microphones exceeds that of conventional 5-microphone arrays.

In the audio modality, state-of-the-art watermarking methods leverage deep neural networks to allow the embedding of human-imperceptible signatures in generated audio. The ideal is to embed signatures that can be detected with high accuracy when the watermarked audio is altered via compression, filtering, or other transformations. Existing audio watermarking techniques operate in a post-hoc manner, manipulating "low-level" features of audio recordings after generation (e.g. through the addition of a low-magnitude watermark signal). We show that this post-hoc formulation makes existing audio watermarks vulnerable to transformation-based removal attacks. Focusing on speech audio, we (1) unify and extend existing evaluations of the effect of audio transformations on watermark detectability, and (2) demonstrate that state-of-the-art post-hoc audio watermarks can be removed with no knowledge of the watermarking scheme and minimal degradation in audio quality.

In mixed reality applications, a realistic acoustic experience in spatial environments is as crucial as the visual experience for achieving true immersion. Despite recent advances in neural approaches for Room Impulse Response (RIR) estimation, most existing methods are limited to the single environment on which they are trained, lacking the ability to generalize to new rooms with different geometries and surface materials. We aim to develop a unified model capable of reconstructing the spatial acoustic experience of any environment with minimum additional measurements. To this end, we present xRIR, a framework for cross-room RIR prediction. The core of our generalizable approach lies in combining a geometric feature extractor, which captures spatial context from panorama depth images, with a RIR encoder that extracts detailed acoustic features from only a few reference RIR samples. To evaluate our method, we introduce ACOUSTICROOMS, a new dataset featuring high-fidelity simulation of over 300,000 RIRs from 260 rooms. Experiments show that our method strongly outperforms a series of baselines. Furthermore, we successfully perform sim-to-real transfer by evaluating our model on four real-world environments, demonstrating the generalizability of our approach and the realism of our dataset.

Recent advancements in audio language models have underscored the pivotal role of audio tokenization, which converts audio signals into discrete tokens, thereby facilitating the application of language model architectures to the audio domain. In this study, we introduce ALMTokenizer, a novel low-bitrate and semantically rich audio codec tokenizer for audio language models. Prior methods, such as Encodec, typically encode individual audio frames into discrete tokens without considering the use of context information across frames. Unlike these methods, we introduce a novel query-based compression strategy to capture holistic information with a set of learnable query tokens by explicitly modeling the context information across frames. This design not only enables the codec model to capture more semantic information but also encodes the audio signal with fewer token sequences. Additionally, to enhance the semantic information in audio codec models, we introduce the following: (1) A masked autoencoder (MAE) loss, (2) Vector quantization based on semantic priors, and (3) An autoregressive (AR) prediction loss. As a result, ALMTokenizer achieves competitive reconstruction performance relative to state-of-the-art approaches while operating at a lower bitrate. Within the same audio language model framework, ALMTokenizer outperforms previous tokenizers in audio understanding and generation tasks.

With the advancement of speech synthesis technology, users have higher expectations for the naturalness and expressiveness of synthesized speech. But previous research ignores the importance of prompt selection. This study proposes a text-to-speech (TTS) framework based on Retrieval-Augmented Generation (RAG) technology, which can dynamically adjust the speech style according to the text content to achieve more natural and vivid communication effects. We have constructed a speech style knowledge database containing high-quality speech samples in various contexts and developed a style matching scheme. This scheme uses embeddings, extracted by Llama, PER-LLM-Embedder,and Moka, to match with samples in the knowledge database, selecting the most appropriate speech style for synthesis. Furthermore, our empirical research validates the effectiveness of the proposed method. Our demo can be viewed at: https://thuhcsi.github.io/icme2025-AutoStyle-TTS

Automating the synthesis of coordinated bimanual piano performances poses significant challenges, particularly in capturing the intricate choreography between the hands while preserving their distinct kinematic signatures. In this paper, we propose a dual-stream neural framework designed to generate synchronized hand gestures for piano playing from audio input, addressing the critical challenge of modeling both hand independence and coordination. Our framework introduces two key innovations: (i) a decoupled diffusion-based generation framework that independently models each hand's motion via dual-noise initialization, sampling distinct latent noise for each while leveraging a shared positional condition, and (ii) a Hand-Coordinated Asymmetric Attention (HCAA) mechanism suppresses symmetric (common-mode) noise to highlight asymmetric hand-specific features, while adaptively enhancing inter-hand coordination during denoising. The system operates hierarchically: it first predicts 3D hand positions from audio features and then generates joint angles through position-aware diffusion models, where parallel denoising streams interact via HCAA. Comprehensive evaluations demonstrate that our framework outperforms existing state-of-the-art methods across multiple metrics.

Speech synthesis technology has brought great convenience, while the widespread usage of realistic deepfake audio has triggered hazards. Malicious adversaries may unauthorizedly collect victims' speeches and clone a similar voice for illegal exploitation (\textit{e.g.}, telecom fraud). However, the existing defense methods cannot effectively prevent deepfake exploitation and are vulnerable to robust training techniques. Therefore, a more effective and robust data protection method is urgently needed. In response, we propose a defensive framework, \textit{\textbf{SafeSpeech}}, which protects the users' audio before uploading by embedding imperceptible perturbations on original speeches to prevent high-quality synthetic speech. In SafeSpeech, we devise a robust and universal proactive protection technique, \textbf{S}peech \textbf{PE}rturbative \textbf{C}oncealment (\textbf{SPEC}), that leverages a surrogate model to generate universally applicable perturbation for generative synthetic models. Moreover, we optimize the human perception of embedded perturbation in terms of time and frequency domains. To evaluate our method comprehensively, we conduct extensive experiments across advanced models and datasets, both subjectively and objectively. Our experimental results demonstrate that SafeSpeech achieves state-of-the-art (SOTA) voice protection effectiveness and transferability and is highly robust against advanced adaptive adversaries. Moreover, SafeSpeech has real-time capability in real-world tests. The source code is available at \href{https://github.com/wxzyd123/SafeSpeech}{https://github.com/wxzyd123/SafeSpeech}.

Recent studies have demonstrated that vision models can effectively learn multimodal audio-image representations when paired. However, the challenge of enabling deep models to learn representations from unpaired modalities remains unresolved. This issue is especially pertinent in scenarios like Federated Learning (FL), where data is often decentralized, heterogeneous, and lacks a reliable guarantee of paired data. Previous attempts tackled this issue through the use of auxiliary pretrained encoders or generative models on local clients, which invariably raise computational cost with increasing number modalities. Unlike these approaches, in this paper, we aim to address the task of unpaired audio and image recognition using \texttt{FSSUAVL}, a single deep model pretrained in FL with self-supervised contrastive learning (SSL). Instead of aligning the audio and image modalities, \texttt{FSSUAVL} jointly discriminates them by projecting them into a common embedding space using contrastive SSL. This extends the utility of \texttt{FSSUAVL} to paired and unpaired audio and image recognition tasks. Our experiments with CNN and ViT demonstrate that \texttt{FSSUAVL} significantly improves performance across various image- and audio-based downstream tasks compared to using separate deep models for each modality. Additionally, \texttt{FSSUAVL}'s capacity to learn multimodal feature representations allows for integrating auxiliary information, if available, to enhance recognition accuracy.

Real-world speech recordings suffer from degradations such as background noise and reverberation. Speech enhancement aims to mitigate these issues by generating clean high-fidelity signals. While recent generative approaches for speech enhancement have shown promising results, they still face two major challenges: (1) content hallucination, where plausible phonemes generated differ from the original utterance; and (2) inconsistency, failing to preserve speaker's identity and paralinguistic features from the input speech. In this work, we introduce DiTSE (Diffusion Transformer for Speech Enhancement), which addresses quality issues of degraded speech in full bandwidth. Our approach employs a latent diffusion transformer model together with robust conditioning features, effectively addressing these challenges while remaining computationally efficient. Experimental results from both subjective and objective evaluations demonstrate that DiTSE achieves state-of-the-art audio quality that, for the first time, matches real studio-quality audio from the DAPS dataset. Furthermore, DiTSE significantly improves the preservation of speaker identity and content fidelity, reducing hallucinations across datasets compared to state-of-the-art enhancers. Audio samples are available at: http://hguimaraes.me/DiTSE

This paper presents AMNet, an Acoustic Model Network designed to improve the performance of Mandarin speech synthesis by incorporating phrase structure annotation and local convolution modules. AMNet builds upon the FastSpeech 2 architecture while addressing the challenge of local context modeling, which is crucial for capturing intricate speech features such as pauses, stress, and intonation. By embedding a phrase structure parser into the model and introducing a local convolution module, AMNet enhances the model's sensitivity to local information. Additionally, AMNet decouples tonal characteristics from phonemes, providing explicit guidance for tone modeling, which improves tone accuracy and pronunciation. Experimental results demonstrate that AMNet outperforms baseline models in subjective and objective evaluations. The proposed model achieves superior Mean Opinion Scores (MOS), lower Mel Cepstral Distortion (MCD), and improved fundamental frequency fitting $F0 (R^2)$, confirming its ability to generate high-quality, natural, and expressive Mandarin speech.