Sabato Marco Siniscalchi

Chao-Han Huck Yang, I-Fan Chen, Andreas Stolcke et al. · gatech

Differential privacy (DP) is one data protection avenue to safeguard user information used for training deep models by imposing noisy distortion on privacy data. Such a noise perturbation often results in a severe performance degradation in automatic speech recognition (ASR) in order to meet a privacy budget $\varepsilon$. Private aggregation of teacher ensemble (PATE) utilizes ensemble probabilities to improve ASR accuracy when dealing with the noise effects controlled by small values of $\varepsilon$. We extend PATE learning to work with dynamic patterns, namely speech utterances, and perform a first experimental demonstration that it prevents acoustic data leakage in ASR training. We evaluate three end-to-end deep models, including LAS, hybrid CTC/attention, and RNN transducer, on the open-source LibriSpeech and TIMIT corpora. PATE learning-enhanced ASR models outperform the benchmark DP-SGD mechanisms, especially under strict DP budgets, giving relative word error rate reductions between 26.2% and 27.5% for an RNN transducer model evaluated with LibriSpeech. We also introduce a DP-preserving ASR solution for pretraining on public speech corpora.

Pin-Jui Ku, Chao-Han Huck Yang, Sabato Marco Siniscalchi et al. · gatech

We propose a multi-dimensional structured state space (S4) approach to speech enhancement. To better capture the spectral dependencies across the frequency axis, we focus on modifying the multi-dimensional S4 layer with whitening transformation to build new small-footprint models that also achieve good performance. We explore several S4-based deep architectures in time (T) and time-frequency (TF) domains. The 2-D S4 layer can be considered a particular convolutional layer with an infinite receptive field although it utilizes fewer parameters than a conventional convolutional layer. Evaluated on the VoiceBank-DEMAND data set, when compared with the conventional U-net model based on convolutional layers, the proposed TF-domain S4-based model is 78.6% smaller in size, yet it still achieves competitive results with a PESQ score of 3.15 with data augmentation. By increasing the model size, we can even reach a PESQ score of 3.18.

Chao-Han Huck Yang, Taejin Park, Yuan Gong et al. · gatech

Given recent advances in generative AI technology, a key question is how large language models (LLMs) can enhance acoustic modeling tasks using text decoding results from a frozen, pretrained automatic speech recognition (ASR) model. To explore new capabilities in language modeling for speech processing, we introduce the generative speech transcription error correction (GenSEC) challenge. This challenge comprises three post-ASR language modeling tasks: (i) post-ASR transcription correction, (ii) speaker tagging, and (iii) emotion recognition. These tasks aim to emulate future LLM-based agents handling voice-based interfaces while remaining accessible to a broad audience by utilizing open pretrained language models or agent-based APIs. We also discuss insights from baseline evaluations, as well as lessons learned for designing future evaluations.

Chao-Han Huck Yang, Bo Li, Yu Zhang et al. · gatech, nvidia

We propose a quantum kernel learning (QKL) framework to address the inherent data sparsity issues often encountered in training large-scare acoustic models in low-resource scenarios. We project acoustic features based on classical-to-quantum feature encoding. Different from existing quantum convolution techniques, we utilize QKL with features in the quantum space to design kernel-based classifiers. Experimental results on challenging spoken command recognition tasks for a few low-resource languages, such as Arabic, Georgian, Chuvash, and Lithuanian, show that the proposed QKL-based hybrid approach attains good improvements over existing classical and quantum solutions.

Chao-Han Huck Yang, Jun Qi, Sabato Marco Siniscalchi et al. · gatech, nvidia

We propose an ensemble learning framework with Poisson sub-sampling to effectively train a collection of teacher models to issue some differential privacy (DP) guarantee for training data. Through boosting under DP, a student model derived from the training data suffers little model degradation from the models trained with no privacy protection. Our proposed solution leverages upon two mechanisms, namely: (i) a privacy budget amplification via Poisson sub-sampling to train a target prediction model that requires less noise to achieve a same level of privacy budget, and (ii) a combination of the sub-sampling technique and an ensemble teacher-student learning framework that introduces DP-preserving noise at the output of the teacher models and transfers DP-preserving properties via noisy labels. Privacy-preserving student models are then trained with the noisy labels to learn the knowledge with DP-protection from the teacher model ensemble. Experimental evidences on spoken command recognition and continuous speech recognition of Mandarin speech show that our proposed framework greatly outperforms existing DP-preserving algorithms in both speech processing tasks.

Qing Wang, Jun Du, Siyuan Zheng et al. · gatech, nvidia

In this paper, we propose two techniques, namely joint modeling and data augmentation, to improve system performances for audio-visual scene classification (AVSC). We employ pre-trained networks trained only on image data sets to extract video embedding; whereas for audio embedding models, we decide to train them from scratch. We explore different neural network architectures for joint modeling to effectively combine the video and audio modalities. Moreover, data augmentation strategies are investigated to increase audio-visual training set size. For the video modality the effectiveness of several operations in RandAugment is verified. An audio-video joint mixup scheme is proposed to further improve AVSC performances. Evaluated on the development set of TAU Urban Audio Visual Scenes 2021, our final system can achieve the best accuracy of 94.2% among all single AVSC systems submitted to DCASE 2021 Task 1b.

Mohammad Adiban, Kalin Stefanov, Sabato Marco Siniscalchi et al. · gatech

We propose a multi-layer variational autoencoder method, we call HR-VQVAE, that learns hierarchical discrete representations of the data. By utilizing a novel objective function, each layer in HR-VQVAE learns a discrete representation of the residual from previous layers through a vector quantized encoder. Furthermore, the representations at each layer are hierarchically linked to those at previous layers. We evaluate our method on the tasks of image reconstruction and generation. Experimental results demonstrate that the discrete representations learned by HR-VQVAE enable the decoder to reconstruct high-quality images with less distortion than the baseline methods, namely VQVAE and VQVAE-2. HR-VQVAE can also generate high-quality and diverse images that outperform state-of-the-art generative models, providing further verification of the efficiency of the learned representations. The hierarchical nature of HR-VQVAE i) reduces the decoding search time, making the method particularly suitable for high-load tasks and ii) allows to increase the codebook size without incurring the codebook collapse problem.

Chen Chen, Yuchen Hu, Chao-Han Huck Yang et al. · gatech

Code-switching (CS) speech refers to the phenomenon of mixing two or more languages within the same sentence. Despite the recent advances in automatic speech recognition (ASR), CS-ASR is still a challenging task ought to the grammatical structure complexity of the phenomenon and the data scarcity of specific training corpus. In this work, we propose to leverage large language models (LLMs) and lists of hypotheses generated by an ASR to address the CS problem. Specifically, we first employ multiple well-trained ASR models for N-best hypotheses generation, with the aim of increasing the diverse and informative elements in the set of hypotheses. Next, we utilize the LLMs to learn the hypotheses-to-transcription (H2T) mapping by adding a trainable low-rank adapter. Such a generative error correction (GER) method directly predicts the accurate transcription according to its expert linguistic knowledge and N-best hypotheses, resulting in a paradigm shift from the traditional language model rescoring or error correction techniques. Experimental evidence demonstrates that GER significantly enhances CS-ASR accuracy, in terms of reduced mixed error rate (MER). Furthermore, LLMs show remarkable data efficiency for H2T learning, providing a potential solution to the data scarcity problem of CS-ASR in low-resource languages.

Nicole Dalia Cilia, Claudio De Stefano, Francesco Fontanella et al. · gatech

Using kinematic properties of handwriting to support the diagnosis of neurodegenerative disease is a real challenge: non-invasive detection techniques combined with machine learning approaches promise big steps forward in this research field. In literature, the tasks proposed focused on different cognitive skills to elicitate handwriting movements. In particular, the meaning and phonology of words to copy can compromise writing fluency. In this paper, we investigated how word semantics and phonology affect the handwriting of people affected by Alzheimer's disease. To this aim, we used the data from six handwriting tasks, each requiring copying a word belonging to one of the following categories: regular (have a predictable phoneme-grapheme correspondence, e.g., cat), non-regular (have atypical phoneme-grapheme correspondence, e.g., laugh), and non-word (non-meaningful pronounceable letter strings that conform to phoneme-grapheme conversion rules). We analyzed the data using a machine learning approach by implementing four well-known and widely-used classifiers and feature selection. The experimental results showed that the feature selection allowed us to derive a different set of highly distinctive features for each word type. Furthermore, non-regular words needed, on average, more features but achieved excellent classification performance: the best result was obtained on a non-regular, reaching an accuracy close to 90%.

Mohammad Adiban, Kalin Stefanov, Sabato Marco Siniscalchi et al. · gatech

We address the video prediction task by putting forth a novel model that combines (i) a novel hierarchical residual learning vector quantized variational autoencoder (HR-VQVAE), and (ii) a novel autoregressive spatiotemporal predictive model (AST-PM). We refer to this approach as a sequential hierarchical residual learning vector quantized variational autoencoder (S-HR-VQVAE). By leveraging the intrinsic capabilities of HR-VQVAE at modeling still images with a parsimonious representation, combined with the AST-PM's ability to handle spatiotemporal information, S-HR-VQVAE can better deal with major challenges in video prediction. These include learning spatiotemporal information, handling high dimensional data, combating blurry prediction, and implicit modeling of physical characteristics. Extensive experimental results on four challenging tasks, namely KTH Human Action, TrafficBJ, Human3.6M, and Kitti, demonstrate that our model compares favorably against state-of-the-art video prediction techniques both in quantitative and qualitative evaluations despite a much smaller model size. Finally, we boost S-HR-VQVAE by proposing a novel training method to jointly estimate the HR-VQVAE and AST-PM parameters.

Tsun-An Hsieh, Chao-Han Huck Yang, Pin-Yu Chen et al. · gatech, nvidia

This study addresses the speech enhancement (SE) task within the causal inference paradigm by modeling the noise presence as an intervention. Based on the potential outcome framework, the proposed causal inference-based speech enhancement (CISE) separates clean and noisy frames in an intervened noisy speech using a noise detector and assigns both sets of frames to two mask-based enhancement modules (EMs) to perform noise-conditional SE. Specifically, we use the presence of noise as guidance for EM selection during training, and the noise detector selects the enhancement module according to the prediction of the presence of noise for each frame. Moreover, we derived a SE-specific average treatment effect to quantify the causal effect adequately. Experimental evidence demonstrates that CISE outperforms a non-causal mask-based SE approach in the studied settings and has better performance and efficiency than more complex SE models.

Chun-wei Ho, Sabato Marco Siniscalchi, Kai Li et al. · gatech

We propose a knowledge-driven, model-based approach to segmenting audio into single-category and mixed-category chunks with applications to source separation. "Knowledge" here denotes information associated with the data, such as music scores. "Model" here refers to tool that can be used for audio segmentation and recognition, such as hidden Markov models. In contrast to conventional learning that often relies on annotated data with given segment categories and their corresponding boundaries to guide the learning process, the proposed framework does not depend on any pre-segmented training data and learns directly from the input audio and its related knowledge sources to build all necessary models autonomously. Evaluation on simulation data shows that score-guided learning achieves very good music segmentation and separation results. Tested on movie track data for cinematic audio source separation also shows that utilizing sound category knowledge achieves better separation results than those obtained with data-driven techniques without using such information.

Ryandhimas E. Zezario, Dyah A. M. G. Wisnu, Szu-Wei Fu et al. · gatech

In this paper, we introduce GatherMOS, a novel framework that leverages large language models (LLM) as meta-evaluators to aggregate diverse signals into quality predictions. GatherMOS integrates lightweight acoustic descriptors with pseudo-labels from DNSMOS and VQScore, enabling the LLM to reason over heterogeneous inputs and infer perceptual mean opinion scores (MOS). We further explore both zero-shot and few-shot in-context learning setups, showing that zero-shot GatherMOS maintains stable performance across diverse conditions, while few-shot guidance yields large gains when support samples match the test conditions. Experiments on the VoiceBank-DEMAND dataset demonstrate that GatherMOS consistently outperforms DNSMOS, VQScore, naive score averaging, and even learning-based models such as CNN-BLSTM and MOS-SSL when trained under limited labeled-data conditions. These results highlight the potential of LLM-based aggregation as a practical strategy for non-intrusive speech quality evaluation.

Moreno La Quatra, Alkis Koudounas, Lorenzo Vaiani et al. · gatech

Limited diversity in standardized benchmarks for evaluating audio representation learning (ARL) methods may hinder systematic comparison of current methods' capabilities. We present ARCH, a comprehensive benchmark for evaluating ARL methods on diverse audio classification domains, covering acoustic events, music, and speech. ARCH comprises 12 datasets, that allow us to thoroughly assess pre-trained SSL models of different sizes. ARCH streamlines benchmarking of ARL techniques through its unified access to a wide range of domains and its ability to readily incorporate new datasets and models. To address the current lack of open-source, pre-trained models for non-speech audio, we also release new pre-trained models that demonstrate strong performance on non-speech datasets. We argue that the presented wide-ranging evaluation provides valuable insights into state-of-the-art ARL methods, and is useful to pinpoint promising research directions.

Hu Hu, Chao-Han Huck Yang, Xianjun Xia et al.

In this technical report, we present a joint effort of four groups, namely GT, USTC, Tencent, and UKE, to tackle Task 1 - Acoustic Scene Classification (ASC) in the DCASE 2020 Challenge. Task 1 comprises two different sub-tasks: (i) Task 1a focuses on ASC of audio signals recorded with multiple (real and simulated) devices into ten different fine-grained classes, and (ii) Task 1b concerns with classification of data into three higher-level classes using low-complexity solutions. For Task 1a, we propose a novel two-stage ASC system leveraging upon ad-hoc score combination of two convolutional neural networks (CNNs), classifying the acoustic input according to three classes, and then ten classes, respectively. Four different CNN-based architectures are explored to implement the two-stage classifiers, and several data augmentation techniques are also investigated. For Task 1b, we leverage upon a quantization method to reduce the complexity of two of our top-accuracy three-classes CNN-based architectures. On Task 1a development data set, an ASC accuracy of 76.9\% is attained using our best single classifier and data augmentation. An accuracy of 81.9\% is then attained by a final model fusion of our two-stage ASC classifiers. On Task 1b development data set, we achieve an accuracy of 96.7\% with a model size smaller than 500KB. Code is available: https://github.com/MihawkHu/DCASE2020_task1.

Jun Qi, Hu Hu, Yannan Wang et al.

We propose a tensor-to-vector regression approach to multi-channel speech enhancement in order to address the issue of input size explosion and hidden-layer size expansion. The key idea is to cast the conventional deep neural network (DNN) based vector-to-vector regression formulation under a tensor-train network (TTN) framework. TTN is a recently emerged solution for compact representation of deep models with fully connected hidden layers. Thus TTN maintains DNN's expressive power yet involves a much smaller amount of trainable parameters. Furthermore, TTN can handle a multi-dimensional tensor input by design, which exactly matches the desired setting in multi-channel speech enhancement. We first provide a theoretical extension from DNN to TTN based regression. Next, we show that TTN can attain speech enhancement quality comparable with that for DNN but with much fewer parameters, e.g., a reduction from 27 million to only 5 million parameters is observed in a single-channel scenario. TTN also improves PESQ over DNN from 2.86 to 2.96 by slightly increasing the number of trainable parameters. Finally, in 8-channel conditions, a PESQ of 3.12 is achieved using 20 million parameters for TTN, whereas a DNN with 68 million parameters can only attain a PESQ of 3.06. Our implementation is available online https://github.com/uwjunqi/Tensor-Train-Neural-Network.

Rong Chao, Wen-Huang Cheng, Moreno La Quatra et al. · gatech

This work aims to investigate the use of a recently proposed, attention-free, scalable state-space model (SSM), Mamba, for the speech enhancement (SE) task. In particular, we employ Mamba to deploy different regression-based SE models (SEMamba) with different configurations, namely basic, advanced, causal, and non-causal. Furthermore, loss functions either based on signal-level distances or metric-oriented are considered. Experimental evidence shows that SEMamba attains a competitive PESQ of 3.55 on the VoiceBank-DEMAND dataset with the advanced, non-causal configuration. A new state-of-the-art PESQ of 3.69 is also reported when SEMamba is combined with Perceptual Contrast Stretching (PCS). Compared against Transformed-based equivalent SE solutions, a noticeable FLOPs reduction up to ~12% is observed with the advanced non-causal configurations. Finally, SEMamba can be used as a pre-processing step before automatic speech recognition (ASR), showing competitive performance against recent SE solutions.

Chen Chen, Ruizhe Li, Yuchen Hu et al. · gatech

Recent studies have successfully shown that large language models (LLMs) can be successfully used for generative error correction (GER) on top of the automatic speech recognition (ASR) output. Specifically, an LLM is utilized to carry out a direct mapping from the N-best hypotheses list generated by an ASR system to the predicted output transcription. However, despite its effectiveness, GER introduces extra data uncertainty since the LLM is trained without taking into account acoustic information available in the speech signal. In this work, we aim to overcome such a limitation by infusing acoustic information before generating the predicted transcription through a novel late fusion solution termed Uncertainty-Aware Dynamic Fusion (UADF). UADF is a multimodal fusion approach implemented into an auto-regressive decoding process and works in two stages: (i) It first analyzes and calibrates the token-level LLM decision, and (ii) it then dynamically assimilates the information from the acoustic modality. Experimental evidence collected from various ASR tasks shows that UADF surpasses existing fusion mechanisms in several ways. It yields significant improvements in word error rate (WER) while mitigating data uncertainty issues in LLM and addressing the poor generalization relied with sole modality during fusion. We also demonstrate that UADF seamlessly adapts to audio-visual speech recognition.

Moreno La Quatra, Valerio Mario Salerno, Yu Tsao et al. · gatech

In this paper, we present an encoder-decoder model leveraging Flan-T5 for post-Automatic Speech Recognition (ASR) Generative Speech Error Correction (GenSEC), and we refer to it as FlanEC. We explore its application within the GenSEC framework to enhance ASR outputs by mapping n-best hypotheses into a single output sentence. By utilizing n-best lists from ASR models, we aim to improve the linguistic correctness, accuracy, and grammaticality of final ASR transcriptions. Specifically, we investigate whether scaling the training data and incorporating diverse datasets can lead to significant improvements in post-ASR error correction. We evaluate FlanEC using the HyPoradise dataset, providing a comprehensive analysis of the model's effectiveness in this domain. Furthermore, we assess the proposed approach under different settings to evaluate model scalability and efficiency, offering valuable insights into the potential of instruction-tuned encoder-decoder models for this task.

Haoyang Li, Yuchen Hu, Chen Chen et al. · gatech

Deep neural network (DNN)-based speech enhancement (SE) usually uses conventional activation functions, which lack the expressiveness to capture complex multiscale structures needed for high-fidelity SE. Group-Rational KAN (GR-KAN), a variant of Kolmogorov-Arnold Networks (KAN), retains KAN's expressiveness while improving scalability on complex tasks. We adapt GR-KAN to existing DNN-based SE by replacing dense layers with GR-KAN layers in the time-frequency (T-F) domain MP-SENet and adapting GR-KAN's activations into the 1D CNN layers in the time-domain Demucs. Results on Voicebank-DEMAND show that GR-KAN requires up to 4x fewer parameters while improving PESQ by up to 0.1. In contrast, KAN, facing scalability issues, outperforms MLP on a small-scale signal modeling task but fails to improve MP-SENet. We demonstrate the first successful use of KAN-based methods for consistent improvement in both time- and SoTA TF-domain SE, establishing GR-KAN as a promising alternative for SE.

Yu-Tung Liu, Kuan-Chen Wang, Rong Chao et al. · gatech

Surface electromyography (sEMG) recordings can be contaminated by electrocardiogram (ECG) signals when the monitored muscle is closed to the heart. Traditional signal processing-based approaches, such as high-pass filtering and template subtraction, have been used to remove ECG interference but are often limited in their effectiveness. Recently, neural network-based methods have shown greater promise for sEMG denoising, but they still struggle to balance both efficiency and effectiveness. In this study, we introduce MSEMG, a novel system that integrates the Mamba state space model with a convolutional neural network to serve as a lightweight sEMG denoising model. We evaluated MSEMG using sEMG data from the Non-Invasive Adaptive Prosthetics database and ECG signals from the MIT-BIH Normal Sinus Rhythm Database. The results show that MSEMG outperforms existing methods, generating higher-quality sEMG signals using fewer parameters.

Whenty Ariyanti, Kuan-Yu Chen, Sabato Marco Siniscalchi et al. · gatech

Perceptual voice quality assessment is essential for diagnosing and monitoring voice disorders by providing standardized evaluations of vocal function. Traditionally, expert raters use standard scales such as the Consensus Auditory-Perceptual Evaluation of Voice (CAPE-V) and Grade, Roughness, Breathiness, Asthenia, and Strain (GRBAS). However, these metrics are subjective and prone to inter-rater variability, motivating the need for automated, objective assessment methods. This study proposes Voice Quality Assessment Network (VOQANet), a deep learning-based framework with an attention mechanism that leverages a Speech Foundation Model (SFM) to extract high-level acoustic and prosodic information from raw speech. To enhance robustness and interpretability, we also introduce VOQANet+, which integrates low-level speech descriptors such as jitter, shimmer, and harmonics-to-noise ratio (HNR) with SFM embeddings into a hybrid representation. Unlike prior studies focused only on vowel-based phonation (PVQD-A subset) of the Perceptual Voice Quality Dataset (PVQD), we evaluate our models on both vowel-based and sentence-level speech (PVQD-S subset) to improve generalizability. Results show that sentence-based input outperforms vowel-based input, especially at the patient level, underscoring the value of longer utterances for capturing perceptual voice attributes. VOQANet consistently surpasses baseline methods in root mean squared error (RMSE) and Pearson correlation coefficient (PCC) across CAPE-V and GRBAS dimensions, with VOQANet+ achieving even better performance. Additional experiments under noisy conditions show that VOQANet+ maintains high prediction accuracy and robustness, supporting its potential for real-world and telehealth deployment.

Alkis Koudounas, Moreno La Quatra, Manuel Giollo et al. · gatech

Hallucinations in automatic speech recognition (ASR) systems refer to fluent and coherent transcriptions produced by neural ASR models that are completely unrelated to the underlying acoustic input (i.e., the speech signal). While similar to conventional decoding errors in potentially compromising the usability of transcriptions for downstream applications, hallucinations can be more detrimental due to their preservation of syntactically and semantically plausible structure. This apparent coherence can mislead subsequent processing stages and introduce serious risks, particularly in critical domains such as healthcare and law. Conventional evaluation metrics are primarily centered on error-based metrics and fail to distinguish between phonetic inaccuracies and hallucinations. Consequently, there is a critical need for new evaluation frameworks that can effectively identify and assess models with a heightened propensity for generating hallucinated content. To this end, we introduce SHALLOW, the first benchmark framework that systematically categorizes and quantifies hallucination phenomena in ASR along four complementary axes: lexical, phonetic, morphological, and semantic. We define targeted metrics within each category to produce interpretable profiles of model behavior. Through evaluation across various architectures and speech domains, we have found that SHALLOW metrics correlate strongly with word error rate (WER) when recognition quality is high (i.e., low WER). Still, this correlation weakens substantially as WER increases. SHALLOW, therefore, captures fine-grained error patterns that WER fails to distinguish under degraded and challenging conditions. Our framework supports specific diagnosis of model weaknesses and provides feedback for model improvement beyond what aggregate error rates can offer.

Tong Guan, Zijie Meng, Dianqi Li et al. · gatech

Recent advances in multimodal time series learning underscore a paradigm shift from analytics centered on basic patterns toward advanced time series understanding and reasoning. However, existing multimodal time series datasets mostly remain at the level of surface alignment and question answering, without reaching the depth of genuine reasoning. The absence of well-defined tasks that genuinely require time series reasoning, along with the scarcity of high-quality data, has limited progress in building practical time series reasoning models (TSRMs). To this end, we introduce Time Series Reasoning Suite (TSR-Suite), which formalizes four atomic tasks that span three fundamental capabilities for reasoning with time series: (1) perception, acquired through scenario understanding and causality discovery; (2) extrapolation, realized via event-aware forecasting; and (3) decision-making, developed through deliberation over perception and extrapolation. TSR-Suite is the first comprehensive time series reasoning suite that supports not only thorough evaluation but also the data pipeline and training of TSRMs. It contains more than 23K samples, of which 2.3K are carefully curated through a human-guided hierarchical annotation process. Building on this foundation, we introduce TimeOmni-1, the first unified reasoning model designed to address diverse real-world problems demanding time series reasoning. The model is trained in multiple stages, integrating a mixture of task scenarios, novel reward functions, and tailored optimizations. Experiments show that TimeOmni-1 delivers strong out-of-distribution generalization across all tasks and achieves a high rate of valid responses. It significantly improves causality discovery accuracy (64.0% vs. 35.9% with GPT-4.1) and raises the valid response rate by over 6% compared to GPT-4.1 on the event-aware forecasting task.

Haoyang Li, Nana Hou, Yuchen Hu et al. · gatech

This work investigates speech enhancement (SE) from the perspective of language models (LMs). We propose a novel method that leverages Direct Preference Optimization (DPO) to improve the perceptual quality of enhanced speech. Using UTMOS, a neural MOS prediction model, as a proxy for human ratings, our approach guides optimization toward perceptually preferred outputs. This differs from existing LM-based SE methods that focus on maximizing the likelihood of clean speech tokens, which may misalign with human perception and degrade quality despite low prediction error. Experiments on the 2020 Deep Noise Suppression Challenge test sets demonstrate that applying DPO to a pretrained LM-based SE model yields consistent improvements across various speech quality metrics, with relative gains of up to 56%. To our knowledge, this is the first application of DPO to SE and the first to incorporate proxy perceptual feedback into LM-based SE training, pointing to a promising direction for perceptually aligned SE.

Jun Qi, Chen-Yu Liu, Sabato Marco Siniscalchi et al. · gatech

Low-Rank Adaptation (LoRA) is widely recognized for its parameter-efficient fine-tuning of large-scale neural models. However, standard LoRA independently optimizes low-rank matrices, which inherently limits its expressivity and generalization capabilities. While classical tensor-train (TT) decomposition can be separately employed on individual LoRA matrices, this work demonstrates that the classical TT-based approach neither significantly improves parameter efficiency nor achieves substantial performance gains. This paper proposes TensorGuide, a novel tensor-train-guided adaptation framework to overcome these limitations. TensorGuide generates two correlated low-rank LoRA matrices through a unified TT structure driven by controlled Gaussian noise. The resulting joint TT representation inherently provides structured, low-rank adaptations, significantly enhancing expressivity, generalization, and parameter efficiency without increasing the number of trainable parameters. Theoretically, we justify these improvements through neural tangent kernel analyses, demonstrating superior optimization dynamics and enhanced generalization. Extensive experiments on quantum dot classification and GPT-2 fine-tuning benchmarks demonstrate that TensorGuide-based LoRA consistently outperforms standard LoRA and TT-LoRA, achieving improved accuracy and scalability with fewer parameters.

Moreno La Quatra, Alkis Koudounas, Elena Baralis et al.

Italy exhibits rich linguistic diversity across its territory due to the distinct regional languages spoken in different areas. Recent advances in self-supervised learning provide new opportunities to analyze Italy's linguistic varieties using speech data alone. This includes the potential to leverage representations learned from large amounts of data to better examine nuances between closely related linguistic varieties. In this study, we focus on automatically identifying the geographic region of origin of speech samples drawn from Italy's diverse language varieties. We leverage self-supervised learning models to tackle this task and analyze differences and similarities between Italy's regional languages. In doing so, we also seek to uncover new insights into the relationships among these diverse yet closely related varieties, which may help linguists understand their interconnected evolution and regional development over time and space. To improve the discriminative ability of learned representations, we evaluate several supervised contrastive learning objectives, both as pre-training steps and additional fine-tuning objectives. Experimental evidence shows that pre-trained self-supervised models can effectively identify regions from speech recording. Additionally, incorporating contrastive objectives during fine-tuning improves classification accuracy and yields embeddings that distinctly separate regional varieties, demonstrating the value of combining self-supervised pre-training and contrastive learning for this task.

Hao Yen, Pin-Jui Ku, Sabato Marco Siniscalchi et al.

We propose a novel language-universal approach to end-to-end automatic spoken keyword recognition (SKR) leveraging upon (i) a self-supervised pre-trained model, and (ii) a set of universal speech attributes (manner and place of articulation). Specifically, Wav2Vec2.0 is used to generate robust speech representations, followed by a linear output layer to produce attribute sequences. A non-trainable pronunciation model then maps sequences of attributes into spoken keywords in a multilingual setting. Experiments on the Multilingual Spoken Words Corpus show comparable performances to character- and phoneme-based SKR in seen languages. The inclusion of domain adversarial training (DAT) improves the proposed framework, outperforming both character- and phoneme-based SKR approaches with 13.73% and 17.22% relative word error rate (WER) reduction in seen languages, and achieves 32.14% and 19.92% WER reduction for unseen languages in zero-shot settings.

Chun-Wei Ho, Chao-Han Huck Yang, Sabato Marco Siniscalchi

In this work, we devise a parameter-efficient solution to bring differential privacy (DP) guarantees into adaptation of a cross-lingual speech classifier. We investigate a new frozen pre-trained adaptation framework for DP-preserving speech modeling without full model fine-tuning. First, we introduce a noisy teacher-student ensemble into a conventional adaptation scheme leveraging a frozen pre-trained acoustic model and attain superior performance than DP-based stochastic gradient descent (DPSGD). Next, we insert residual adapters (RA) between layers of the frozen pre-trained acoustic model. The RAs reduce training cost and time significantly with a negligible performance drop. Evaluated on the open-access Multilingual Spoken Words (MLSW) dataset, our solution reduces the number of trainable parameters by 97.5% using the RAs with only a 4% performance drop with respect to fine-tuning the cross-lingual speech classifier while preserving DP guarantees.

Hu Hu, Sabato Marco Siniscalchi, Chao-Han Huck Yang et al.

We propose a variational Bayesian (VB) approach to learning distributions of latent variables in deep neural network (DNN) models for cross-domain knowledge transfer, to address acoustic mismatches between training and testing conditions. Instead of carrying out point estimation in conventional maximum a posteriori estimation with a risk of having a curse of dimensionality in estimating a huge number of model parameters, we focus our attention on estimating a manageable number of latent variables of DNNs via a VB inference framework. To accomplish model transfer, knowledge learnt from a source domain is encoded in prior distributions of latent variables and optimally combined, in a Bayesian sense, with a small set of adaptation data from a target domain to approximate the corresponding posterior distributions. Experimental results on device adaptation in acoustic scene classification show that our proposed VB approach can obtain good improvements on target devices, and consistently outperforms 13 state-of-the-art knowledge transfer algorithms.

Hao Yen, Pin-Jui Ku, Chao-Han Huck Yang et al.

In this study, we propose a novel adversarial reprogramming (AR) approach for low-resource spoken command recognition (SCR), and build an AR-SCR system. The AR procedure aims to modify the acoustic signals (from the target domain) to repurpose a pretrained SCR model (from the source domain). To solve the label mismatches between source and target domains, and further improve the stability of AR, we propose a novel similarity-based label mapping technique to align classes. In addition, the transfer learning (TL) technique is combined with the original AR process to improve the model adaptation capability. We evaluate the proposed AR-SCR system on three low-resource SCR datasets, including Arabic, Lithuanian, and dysarthric Mandarin speech. Experimental results show that with a pretrained AM trained on a large-scale English dataset, the proposed AR-SCR system outperforms the current state-of-the-art results on Arabic and Lithuanian speech commands datasets, with only a limited amount of training data.

Zhen Huang, Xiaodan Zhuang, Daben Liu et al.

In this paper, we present our initial efforts for building a code-switching (CS) speech recognition system leveraging existing acoustic models (AMs) and language models (LMs), i.e., no training required, and specifically targeting intra-sentential switching. To achieve such an ambitious goal, new mechanisms for foreign pronunciation generation and language model (LM) enrichment have been devised. Specifically, we have designed an automatic approach to obtain high quality pronunciation of foreign language (FL) words in the native language (NL) phoneme set using existing acoustic phone decoders and an LSTM-based grapheme-to-phoneme (G2P) model. Improved accented pronunciations have thus been obtained by learning foreign pronunciations directly from data. Furthermore, a code-switching LM was deployed by converting the original NL LM into a CS LM using translated word pairs and borrowing statistics for the NL LM. Experimental evidence clearly demonstrates that our approach better deals with accented foreign pronunciations than techniques based on human labeling. Moreover, our best system achieves a 55.5% relative word error rate reduction from 34.4%, obtained with a conventional monolingual ASR system, to 15.3% on an intra-sentential CS task without harming the monolingual recognition accuracy.

Hao Yen, Chao-Han Huck Yang, Hu Hu et al.

We propose a novel neural model compression strategy combining data augmentation, knowledge transfer, pruning, and quantization for device-robust acoustic scene classification (ASC). Specifically, we tackle the ASC task in a low-resource environment leveraging a recently proposed advanced neural network pruning mechanism, namely Lottery Ticket Hypothesis (LTH), to find a sub-network neural model associated with a small amount non-zero model parameters. The effectiveness of LTH for low-complexity acoustic modeling is assessed by investigating various data augmentation and compression schemes, and we report an efficient joint framework for low-complexity multi-device ASC, called \emph{Acoustic Lottery}. Acoustic Lottery could compress an ASC model up to $1/10^{4}$ and attain a superior performance (validation accuracy of 79.4% and Log loss of 0.64) compared to its not compressed seed model. All results reported in this work are based on a joint effort of four groups, namely GT-USTC-UKE-Tencent, aiming to address the "Low-Complexity Acoustic Scene Classification (ASC) with Multiple Devices" in the DCASE 2021 Challenge Task 1a.

Chao-Han Huck Yang, Sabato Marco Siniscalchi, Chin-Hui Lee

We propose using an adversarial autoencoder (AAE) to replace generative adversarial network (GAN) in the private aggregation of teacher ensembles (PATE), a solution for ensuring differential privacy in speech applications. The AAE architecture allows us to obtain good synthetic speech leveraging upon a discriminative training of latent vectors. Such synthetic speech is used to build a privacy-preserving classifier when non-sensitive data is not sufficiently available in the public domain. This classifier follows the PATE scheme that uses an ensemble of noisy outputs to label the synthetic samples and guarantee $\varepsilon$-differential privacy (DP) on its derived classifiers. Our proposed framework thus consists of an AAE-based generator and a PATE-based classifier (PATE-AAE). Evaluated on the Google Speech Commands Dataset Version II, the proposed PATE-AAE improves the average classification accuracy by +$2.11\%$ and +$6.60\%$, respectively, when compared with alternative privacy-preserving solutions, namely PATE-GAN and DP-GAN, while maintaining a strong level of privacy target at $\varepsilon$=0.01 with a fixed $δ$=10$^{-5}$.

Hu Hu, Chao-Han Huck Yang, Xianjun Xia et al.

To improve device robustness, a highly desirable key feature of a competitive data-driven acoustic scene classification (ASC) system, a novel two-stage system based on fully convolutional neural networks (CNNs) is proposed. Our two-stage system leverages on an ad-hoc score combination based on two CNN classifiers: (i) the first CNN classifies acoustic inputs into one of three broad classes, and (ii) the second CNN classifies the same inputs into one of ten finer-grained classes. Three different CNN architectures are explored to implement the two-stage classifiers, and a frequency sub-sampling scheme is investigated. Moreover, novel data augmentation schemes for ASC are also investigated. Evaluated on DCASE 2020 Task 1a, our results show that the proposed ASC system attains a state-of-the-art accuracy on the development set, where our best system, a two-stage fusion of CNN ensembles, delivers a 81.9% average accuracy among multi-device test data, and it obtains a significant improvement on unseen devices. Finally, neural saliency analysis with class activation mapping (CAM) gives new insights on the patterns learnt by our models.

Chao-Han Huck Yang, Jun Qi, Samuel Yen-Chi Chen et al.

We propose a novel decentralized feature extraction approach in federated learning to address privacy-preservation issues for speech recognition. It is built upon a quantum convolutional neural network (QCNN) composed of a quantum circuit encoder for feature extraction, and a recurrent neural network (RNN) based end-to-end acoustic model (AM). To enhance model parameter protection in a decentralized architecture, an input speech is first up-streamed to a quantum computing server to extract Mel-spectrogram, and the corresponding convolutional features are encoded using a quantum circuit algorithm with random parameters. The encoded features are then down-streamed to the local RNN model for the final recognition. The proposed decentralized framework takes advantage of the quantum learning progress to secure models and to avoid privacy leakage attacks. Testing on the Google Speech Commands Dataset, the proposed QCNN encoder attains a competitive accuracy of 95.12% in a decentralized model, which is better than the previous architectures using centralized RNN models with convolutional features. We also conduct an in-depth study of different quantum circuit encoder architectures to provide insights into designing QCNN-based feature extractors. Neural saliency analyses demonstrate a correlation between the proposed QCNN features, class activation maps, and input spectrograms. We provide an implementation for future studies.

Jun Qi, Jun Du, Sabato Marco Siniscalchi et al.

In this paper, we exploit the properties of mean absolute error (MAE) as a loss function for the deep neural network (DNN) based vector-to-vector regression. The goal of this work is two-fold: (i) presenting performance bounds of MAE, and (ii) demonstrating new properties of MAE that make it more appropriate than mean squared error (MSE) as a loss function for DNN based vector-to-vector regression. First, we show that a generalized upper-bound for DNN-based vector- to-vector regression can be ensured by leveraging the known Lipschitz continuity property of MAE. Next, we derive a new generalized upper bound in the presence of additive noise. Finally, in contrast to conventional MSE commonly adopted to approximate Gaussian errors for regression, we show that MAE can be interpreted as an error modeled by Laplacian distribution. Speech enhancement experiments are conducted to corroborate our proposed theorems and validate the performance advantages of MAE over MSE for DNN based regression.

Jun Qi, Jun Du, Sabato Marco Siniscalchi et al.

In this paper, we show that, in vector-to-vector regression utilizing deep neural networks (DNNs), a generalized loss of mean absolute error (MAE) between the predicted and expected feature vectors is upper bounded by the sum of an approximation error, an estimation error, and an optimization error. Leveraging upon error decomposition techniques in statistical learning theory and non-convex optimization theory, we derive upper bounds for each of the three aforementioned errors and impose necessary constraints on DNN models. Moreover, we assess our theoretical results through a set of image de-noising and speech enhancement experiments. Our proposed upper bounds of MAE for DNN based vector-to-vector regression are corroborated by the experimental results and the upper bounds are valid with and without the "over-parametrization" technique.

Hu Hu, Sabato Marco Siniscalchi, Yannan Wang et al.

In this paper, we propose a domain adaptation framework to address the device mismatch issue in acoustic scene classification leveraging upon neural label embedding (NLE) and relational teacher student learning (RTSL). Taking into account the structural relationships between acoustic scene classes, our proposed framework captures such relationships which are intrinsically device-independent. In the training stage, transferable knowledge is condensed in NLE from the source domain. Next in the adaptation stage, a novel RTSL strategy is adopted to learn adapted target models without using paired source-target data often required in conventional teacher student learning. The proposed framework is evaluated on the DCASE 2018 Task1b data set. Experimental results based on AlexNet-L deep classification models confirm the effectiveness of our proposed approach for mismatch situations. NLE-alone adaptation compares favourably with the conventional device adaptation and teacher student based adaptation techniques. NLE with RTSL further improves the classification accuracy.

Hu Hu, Sabato Marco Siniscalchi, Yannan Wang et al.

In this paper, we propose a sub-utterance unit selection framework to remove acoustic segments in audio recordings that carry little information for acoustic scene classification (ASC). Our approach is built upon a universal set of acoustic segment units covering the overall acoustic scene space. First, those units are modeled with acoustic segment models (ASMs) used to tokenize acoustic scene utterances into sequences of acoustic segment units. Next, paralleling the idea of stop words in information retrieval, stop ASMs are automatically detected. Finally, acoustic segments associated with the stop ASMs are blocked, because of their low indexing power in retrieval of most acoustic scenes. In contrast to building scene models with whole utterances, the ASM-removed sub-utterances, i.e., acoustic utterances without stop acoustic segments, are then used as inputs to the AlexNet-L back-end for final classification. On the DCASE 2018 dataset, scene classification accuracy increases from 68%, with whole utterances, to 72.1%, with segment selection. This represents a competitive accuracy without any data augmentation, and/or ensemble strategy. Moreover, our approach compares favourably to AlexNet-L with attention.

Jun Qi, Hu Hu, Yannan Wang et al.

This paper investigates different trade-offs between the number of model parameters and enhanced speech qualities by employing several deep tensor-to-vector regression models for speech enhancement. We find that a hybrid architecture, namely CNN-TT, is capable of maintaining a good quality performance with a reduced model parameter size. CNN-TT is composed of several convolutional layers at the bottom for feature extraction to improve speech quality and a tensor-train (TT) output layer on the top to reduce model parameters. We first derive a new upper bound on the generalization power of the convolutional neural network (CNN) based vector-to-vector regression models. Then, we provide experimental evidence on the Edinburgh noisy speech corpus to demonstrate that, in single-channel speech enhancement, CNN outperforms DNN at the expense of a small increment of model sizes. Besides, CNN-TT slightly outperforms the CNN counterpart by utilizing only 32\% of the CNN model parameters. Besides, further performance improvement can be attained if the number of CNN-TT parameters is increased to 44\% of the CNN model size. Finally, our experiments of multi-channel speech enhancement on a simulated noisy WSJ0 corpus demonstrate that our proposed hybrid CNN-TT architecture achieves better results than both DNN and CNN models in terms of better-enhanced speech qualities and smaller parameter sizes.

Zhen Huang, Sabato Marco Siniscalchi, I-Fan Chen et al.

We present a Bayesian approach to adapting parameters of a well-trained context-dependent, deep-neural-network, hidden Markov model (CD-DNN-HMM) to improve automatic speech recognition performance. Given an abundance of DNN parameters but with only a limited amount of data, the effectiveness of the adapted DNN model can often be compromised. We formulate maximum a posteriori (MAP) adaptation of parameters of a specially designed CD-DNN-HMM with an augmented linear hidden networks connected to the output tied states, or senones, and compare it to feature space MAP linear regression previously proposed. Experimental evidences on the 20,000-word open vocabulary Wall Street Journal task demonstrate the feasibility of the proposed framework. In supervised adaptation, the proposed MAP adaptation approach provides more than 10% relative error reduction and consistently outperforms the conventional transformation based methods. Furthermore, we present an initial attempt to generate hierarchical priors to improve adaptation efficiency and effectiveness with limited adaptation data by exploiting similarities among senones.