Learning from Random Subspace Exploration: Generalized Test-Time Augmentation with Self-supervised Distillation

We introduce Generalized Test-Time Augmentation (GTTA), a highly effective method for improving the performance of a trained model, which unlike other existing Test-Time Augmentation approaches from the literature is general enough to be used off-the-shelf for many vision and non-vision tasks, such as classification, regression, image segmentation and object detection. By applying a new general data transformation, that randomly perturbs multiple times the PCA subspace projection of a test input, GTTA forms robust ensembles at test time in which, due to sound statistical properties, the structural and systematic noises in the initial input data is filtered out and final estimator errors are reduced. Different from other existing methods, we also propose a final self-supervised learning stage in which the ensemble output, acting as an unsupervised teacher, is used to train the initial single student model, thus reducing significantly the test time computational cost, at no loss in accuracy. Our tests and comparisons to strong TTA approaches and SoTA models on various vision and non-vision well-known datasets and tasks, such as image classification and segmentation, speech recognition and house price prediction, validate the generality of the proposed GTTA. Furthermore, we also prove its effectiveness on the more specific real-world task of salmon segmentation and detection in low-visibility underwater videos, for which we introduce DeepSalmon, the largest dataset of its kind in the literature.

PERTINENCE: Input-based Opportunistic Neural Network Dynamic Execution

Deep neural networks (DNNs) have become ubiquitous thanks to their remarkable ability to model complex patterns across various domains such as computer vision, speech recognition, robotics, etc. While large DNN models are often more accurate than simpler, lightweight models, they are also resource- and energy-hungry. Hence, it is imperative to design methods to reduce reliance on such large models without significant degradation in output accuracy. The high computational cost of these models is often necessary only for a reduced set of challenging inputs, while lighter models can handle most simple ones. Thus, carefully combining properties of existing DNN models in a dynamic, input-based way opens opportunities to improve efficiency without impacting accuracy. In this work, we introduce PERTINENCE, a novel online method designed to analyze the complexity of input features and dynamically select the most suitable model from a pre-trained set to process a given input effectively. To achieve this, we employ a genetic algorithm to explore the training space of an ML-based input dispatcher, enabling convergence towards the Pareto front in the solution space that balances overall accuracy and computational efficiency. We showcase our approach on state-of-the-art Convolutional Neural Networks (CNNs) trained on the CIFAR-10 and CIFAR-100, as well as Vision Transformers (ViTs) trained on TinyImageNet dataset. We report results showing PERTINENCE's ability to provide alternative solutions to existing state-of-the-art models in terms of trade-offs between accuracy and number of operations. By opportunistically selecting among models trained for the same task, PERTINENCE achieves better or comparable accuracy with up to 36% fewer operations.

MambAttention: Mamba with Multi-Head Attention for Generalizable Single-Channel Speech Enhancement

With the advent of new sequence models like Mamba and xLSTM, several studies have shown that these models match or outperform state-of-the-art models in single-channel speech enhancement, automatic speech recognition, and self-supervised audio representation learning. However, prior research has demonstrated that sequence models like LSTM and Mamba tend to overfit to the training set. To address this issue, previous works have shown that adding self-attention to LSTMs substantially improves generalization performance for single-channel speech enhancement. Nevertheless, neither the concept of hybrid Mamba and time-frequency attention models nor their generalization performance have been explored for speech enhancement. In this paper, we propose a novel hybrid architecture, MambAttention, which combines Mamba and shared time- and frequency-multi-head attention modules for generalizable single-channel speech enhancement. To train our model, we introduce VoiceBank+Demand Extended (VB-DemandEx), a dataset inspired by VoiceBank+Demand but with more challenging noise types and lower signal-to-noise ratios. Trained on VB-DemandEx, our proposed MambAttention model significantly outperforms existing state-of-the-art LSTM-, xLSTM-, Mamba-, and Conformer-based systems of similar complexity across all reported metrics on two out-of-domain datasets: DNS 2020 and EARS-WHAM_v2, while matching their performance on the in-domain dataset VB-DemandEx. Ablation studies highlight the role of weight sharing between the time- and frequency-multi-head attention modules for generalization performance. Finally, we explore integrating the shared time- and frequency-multi-head attention modules with LSTM and xLSTM, which yields a notable performance improvement on the out-of-domain datasets. However, our MambAttention model remains superior on both out-of-domain datasets across all reported evaluation metrics.

Accurate, fast, cheap: Choose three. Replacing Multi-Head-Attention with Bidirectional Recurrent Attention for Long-Form ASR

Long-form speech recognition is an application area of increasing research focus. ASR models based on multi-head attention (MHA) are ill-suited to long-form ASR because of their quadratic complexity in sequence length. We build on recent work that has investigated linear complexity recurrent attention (RA) layers for ASR. We find that bidirectional RA layers can match the accuracy of MHA for both short- and long-form applications. We present a strong limited-context attention (LCA) baseline, and show that RA layers are just as accurate while being more efficient. We develop a long-form training paradigm which further improves RA performance, leading to better accuracy than LCA with 44% higher throughput. We also present Direction Dropout, a novel regularization method that improves accuracy, provides fine-grained control of the accuracy/throughput trade-off of bidirectional RA, and enables a new alternating directions decoding mode with even higher throughput.

Hybrid Deep Learning and Signal Processing for Arabic Dialect Recognition in Low-Resource Settings

Arabic dialect recognition presents a significant challenge in speech technology due to the linguistic diversity of Arabic and the scarcity of large annotated datasets, particularly for underrepresented dialects. This research investigates hybrid modeling strategies that integrate classical signal processing techniques with deep learning architectures to address this problem in low-resource scenarios. Two hybrid models were developed and evaluated: (1) Mel-Frequency Cepstral Coefficients (MFCC) combined with a Convolutional Neural Network (CNN), and (2) Discrete Wavelet Transform (DWT) features combined with a Recurrent Neural Network (RNN). The models were trained on a dialect-filtered subset of the Common Voice Arabic dataset, with dialect labels assigned based on speaker metadata. Experimental results demonstrate that the MFCC + CNN architecture achieved superior performance, with an accuracy of 91.2% and strong precision, recall, and F1-scores, significantly outperforming the Wavelet + RNN configuration, which achieved an accuracy of 66.5%. These findings highlight the effectiveness of leveraging spectral features with convolutional models for Arabic dialect recognition, especially when working with limited labeled data. The study also identifies limitations related to dataset size, potential regional overlaps in labeling, and model optimization, providing a roadmap for future research. Recommendations for further improvement include the adoption of larger annotated corpora, integration of self-supervised learning techniques, and exploration of advanced neural architectures such as Transformers. Overall, this research establishes a strong baseline for future developments in Arabic dialect recognition within resource-constrained environments.

Thinking in Directivity: Speech Large Language Model for Multi-Talker Directional Speech Recognition

Recent studies have demonstrated that prompting large language models (LLM) with audio encodings enables effective speech recognition capabilities. However, the ability of Speech LLMs to comprehend and process multi-channel audio with spatial cues remains a relatively uninvestigated area of research. In this work, we present directional-SpeechLlama, a novel approach that leverages the microphone array of smart glasses to achieve directional speech recognition, source localization, and bystander cross-talk suppression. To enhance the model's ability to understand directivity, we propose two key techniques: serialized directional output training (S-DOT) and contrastive direction data augmentation (CDDA). Experimental results show that our proposed directional-SpeechLlama effectively captures the relationship between textual cues and spatial audio, yielding strong performance in both speech recognition and source localization tasks.

An accurate and revised version of optical character recognition-based speech synthesis using LabVIEW

Knowledge extraction through sound is a distinctive property. Visually impaired individuals often rely solely on Braille books and audio recordings provided by NGOs. Due to limitations in these approaches, blind individuals often cannot access books of their choice. Speech is a more effective mode of communication than text for blind and visually impaired persons, as they can easily respond to sounds. This paper presents the development of an accurate, reliable, cost-effective, and user-friendly optical character recognition (OCR)-based speech synthesis system. The OCR-based system has been implemented using Laboratory Virtual Instrument Engineering Workbench (LabVIEW).

Improving Practical Aspects of End-to-End Multi-Talker Speech Recognition for Online and Offline Scenarios

We extend the frameworks of Serialized Output Training (SOT) to address practical needs of both streaming and offline automatic speech recognition (ASR) applications. Our approach focuses on balancing latency and accuracy, catering to real-time captioning and summarization requirements. We propose several key improvements: (1) Leveraging Continuous Speech Separation (CSS) single-channel front-end with end-to-end (E2E) systems for highly overlapping scenarios, challenging the conventional wisdom of E2E versus cascaded setups. The CSS framework improves the accuracy of the ASR system by separating overlapped speech from multiple speakers. (2) Implementing dual models -- Conformer Transducer for streaming and Sequence-to-Sequence for offline -- or alternatively, a two-pass model based on cascaded encoders. (3) Exploring segment-based SOT (segSOT) which is better suited for offline scenarios while also enhancing readability of multi-talker transcriptions.

Exploiting Music Source Separation for Automatic Lyrics Transcription with Whisper

Automatic lyrics transcription (ALT) remains a challenging task in the field of music information retrieval, despite great advances in automatic speech recognition (ASR) brought about by transformer-based architectures in recent years. One of the major challenges in ALT is the high amplitude of interfering audio signals relative to conventional ASR due to musical accompaniment. Recent advances in music source separation have enabled automatic extraction of high-quality separated vocals, which could potentially improve ALT performance. However, the effect of source separation has not been systematically investigated in order to establish best practices for its use. This work examines the impact of source separation on ALT using Whisper, a state-of-the-art open source ASR model. We evaluate Whisper's performance on original audio, separated vocals, and vocal stems across short-form and long-form transcription tasks. For short-form, we suggest a concatenation method that results in a consistent reduction in Word Error Rate (WER). For long-form, we propose an algorithm using source separation as a vocal activity detector to derive segment boundaries, which results in a consistent reduction in WER relative to Whisper's native long-form algorithm. Our approach achieves state-of-the-art results for an open source system on the Jam-ALT long-form ALT benchmark, without any training or fine-tuning. We also publish MUSDB-ALT, the first dataset of long-form lyric transcripts following the Jam-ALT guidelines for which vocal stems are publicly available.

Qwen vs. Gemma Integration with Whisper: A Comparative Study in Multilingual SpeechLLM Systems

This paper presents our system for the MLC-SLM Challenge 2025, focusing on multilingual speech recognition and language modeling with large language models (LLMs). Our approach combines a fine-tuned Whisper-large-v3 encoder with efficient projector architectures and various decoder configurations. We employ a three-stage training methodology that progressively optimizes the encoder, projector, and LLM components. Our system achieves competitive performance with a private test average WER/CER result of 16.63% using the Gemma3-12B and 18.6% using the Qwen2.5-7B as decoder-only language model.