AINeutralarXiv – CS AI · Jun 236/10
🧠Researchers developed multiple approaches to detect hallucinations in OpenAI's Whisper ASR model, where the system generates fluent but unfounded transcriptions. The study found that probing the model's internal decoder states outperformed text-based and LLM-based detection methods, with a hybrid approach combining text metrics and internal representations achieving the best overall performance.
AINeutralarXiv – CS AI · Jun 236/10
🧠Researchers have developed explainable AI techniques to improve trust and understanding of automatic speech recognition (ASR) systems by identifying minimal subsets of audio frames that cause specific transcriptions. The study adapts established XAI methods from image classification and evaluates them against multiple ASR systems including Google API and DeepSpeech using 100 audio samples.
AIBullisharXiv – CS AI · Jun 236/10
🧠Researchers successfully fine-tuned automatic speech recognition (ASR) models to create text corpora for low-resource African languages Fongbe and Hausa, achieving significant improvements in transcription accuracy. The work demonstrates ASR's potential for rapidly expanding language resources in underrepresented languages, though quality varies by linguistic complexity, with Hausa transcriptions approaching production-ready standards while Fongbe requires further refinement.
AINeutralarXiv – CS AI · Jun 236/10
🧠Researchers released WASIL, a dataset of 8,529 Arabic spoken interactions with LLMs including audio, transcriptions, and user feedback, to address how speech recognition errors degrade voice assistant performance. The dataset includes a 2,000-turn test set covering Modern Standard Arabic and four dialects, with annotations distinguishing between genuine unanswerability and ASR-induced failures, enabling more accurate evaluation of voice AI systems.
AINeutralarXiv – CS AI · Jun 236/10
🧠ASTRA is an AI-powered air traffic control training simulator that automates the role of simpilots (human trainers) through advanced speech recognition and response generation systems. The system reduces speech recognition error rates from 107.80% to 23.45% for Singaporean-accented aviation speech and incorporates AI-assisted performance evaluation, addressing a critical training capacity bottleneck in aviation safety infrastructure.
AIBullisharXiv – CS AI · Jun 236/10
🧠Researchers propose Cross-lingual Retrieval-Augmented Classification (CRAC), an AI method that improves dysarthria severity assessment by leveraging speech data from different languages to overcome the scarcity of labeled pathological speech datasets. The approach achieves significant accuracy improvements on Korean and Italian datasets, demonstrating the potential of cross-lingual transfer learning in medical speech analysis.
AINeutralarXiv – CS AI · Jun 236/10
🧠Researchers present a systematic framework for optimizing speech processing models by analyzing tradeoffs between model size, input length, and representation resolution under fixed computational budgets. The study demonstrates non-linear scaling behavior, showing diminishing returns from model scaling and identifying practical efficiency gains through token resolution reduction without significant performance degradation.
AINeutralarXiv – CS AI · Jun 235/10
🧠Researchers conducted a comprehensive layer-wise analysis of how four major self-supervised learning (SSL) speech models encode age and gender information in children's speech. The study reveals that age and gender cues are unevenly distributed across model layers, with early-to-mid layers capturing the strongest paralinguistic signals, and demonstrates reliable classification accuracy even from 1-3 second audio segments.
AINeutralarXiv – CS AI · Jun 196/10
🧠Researchers have developed improved acoustic modeling techniques for recognizing dysarthric speech in children, achieving 4.65% relative improvement in word recognition and 4.63% in sentence recognition using Factorized Time Delay Neural Networks. The study demonstrates that strategic selection of acoustic features, particularly pitch characteristics, significantly enhances performance on low-resource speech recognition tasks.
AINeutralarXiv – CS AI · Jun 196/10
🧠Researchers have achieved significant improvements in dysarthric speech recognition by systematically combining acoustic features with the Factorized Time Delay Neural Network (F-TDNN) model, demonstrating 4.65% relative improvement in word recognition and 4.63% in sentence recognition. The study identifies pitch features as particularly effective for handling the acoustic variability characteristic of impaired speech, advancing accessibility technology for individuals with speech disorders.
AINeutralarXiv – CS AI · Jun 196/10
🧠Researchers developed data augmentation techniques to improve automatic speech recognition (ASR) for people with dysarthria by fine-tuning the Wav2Vec2 model. Using methods like speaking-rate modification, pitch modification, and formant modification tailored to different severity levels, the study achieved significant word error rate reductions across low, medium, and high severity dysarthric speech.
AINeutralarXiv – CS AI · Jun 196/10
🧠Researchers propose a code-mixing guided synthetic speech generation framework to improve automatic speech recognition (ASR) for multilingual code-switching scenarios. By optimizing synthetic data generation using the Code Mixing Index metric, the method demonstrates significant error rate reductions on Mandarin-English speech datasets, addressing a critical limitation in training data availability for code-switched ASR systems.
AINeutralarXiv – CS AI · Jun 195/10
🧠Researchers developed improved Automatic Speech Recognition (ASR) models for Quranic recitation using pretrained Transformer architectures (Wav2Vec2.0, HuBERT, XLS-R), achieving 8% word error rates compared to 16.3% baseline performance. The study demonstrates that domain-specific fine-tuning with 870+ hours of professional and user-recited Quranic audio, combined with Arabic text without diacritics, significantly enhances transcription accuracy while reducing training time by 71%.
AIBullisharXiv – CS AI · Jun 116/10
🧠Researchers demonstrate that pretrained self-supervised speech models (Wav2Vec2 and HuBERT) can accurately recognize click consonants from low-resource Khoisan languages despite training data heavily skewed toward high-resource languages. Fine-tuning on click-rich language data reveals these models generalize better to rare phonemes than expected, suggesting self-supervision creates robust representations across diverse human speech sounds.
AINeutralarXiv – CS AI · Jun 106/10
🧠Researchers have developed an automated system for evaluating Korean toddler pronunciation using speaker diarization and self-supervised learning models, addressing a significant gap in speech assessment tools for this demographic. The system achieved balanced accuracies of 0.720 for consonants and 0.845 for vowels by routing predictions through specialized SSL models, offering potential clinical applications for detecting speech sound disorders affecting nearly half of Korean pediatric cases.
AINeutralHugging Face Blog · Jun 96/10
🧠Researchers benchmark frontier automatic speech recognition (ASR) systems on code-switched speech, where bilingual speakers mix languages mid-conversation. The study evaluates how well modern voice AI handles this common real-world scenario, revealing performance gaps that matter for customer service applications.
AINeutralarXiv – CS AI · Jun 96/10
🧠Researchers present a rigorous study of fine-tuning OpenAI's Whisper model for Swiss German speech recognition, achieving 25.6% WER with honest evaluation on disjoint test data. The work exposes significant benchmark contamination in published Swiss German ASR results, revealing that previous state-of-the-art claims were inflated by models memorizing test sets rather than genuinely understanding dialect.
🏢 OpenAI🏢 Nvidia
AINeutralarXiv – CS AI · Jun 95/10
🧠Researchers propose a training-free method for improving automatic speech recognition in noisy environments by intelligently fusing noisy and speech-enhanced audio based on intelligibility estimates. The approach eliminates the need for trained neural predictors, reducing complexity while maintaining robustness across diverse speech enhancement and ASR model combinations.
AINeutralarXiv – CS AI · Jun 86/10
🧠Researchers propose IRAF, a lightweight module that improves full-duplex spoken dialogue systems by filtering interference from background speakers. The technology uses adaptive fusion to modulate user audio reliability frame-by-frame, demonstrating improved response quality and stable turn-taking in noisy acoustic environments.
AINeutralarXiv – CS AI · Jun 85/10
🧠Researchers achieved state-of-the-art performance on raw waveform acoustic models for phone recognition using CNN-LSTM architectures, with error rates of 13.9%/15.3% on TIMIT benchmarks. Analysis reveals that different phonetic classes benefit differently from model components, and transfer learning from WSJ data improves consonant recognition significantly more than vowels.
AIBullishHugging Face Blog · Jun 46/10
🧠This article provides guidance on fine-tuning Nemotron 3.5 ASR, NVIDIA's automatic speech recognition model, to improve accuracy for specific languages, domains, and accents. The tutorial enables developers to customize the open-source model for specialized use cases beyond its default training data.
AIBullisharXiv – CS AI · Jun 26/10
🧠Researchers introduce a layer-wise projection mapping technique for knowledge distillation that enables efficient model compression, reducing trainable parameters to under 1% of the teacher model while maintaining performance improvements. Combined with LoRA injection, this approach significantly outperforms traditional distillation methods in word error rate metrics and enables rapid parallel training without the computational overhead of mixture-of-experts models.
AIBullisharXiv – CS AI · Jun 26/10
🧠Researchers introduce Murmur, an inference system that optimizes long-form automatic speech recognition by balancing accuracy and latency through a two-level approach: intermediate chunk sizes at the inter-chunk level and attention sparsity exploitation at the intra-chunk level. The system achieves 4.2x latency reduction while maintaining single-pass accuracy on benchmark tests.
AINeutralarXiv – CS AI · Jun 26/10
🧠Echo is a proof-of-concept audio system that unifies speaker diarization, speech recognition, and source separation on a single 25M-parameter ViT encoder pretrained with joint-embedding predictive architecture (JEPA). The system demonstrates competitive performance across three tasks simultaneously without per-task fine-tuning, though it represents a design exploration rather than state-of-the-art on individual metrics.
AINeutralarXiv – CS AI · Jun 26/10
🧠Researchers introduce VocSim, a training-free benchmark for evaluating audio embeddings' ability to identify content across diverse sound sources without parameter updates or labeled data. Testing 125k clips spanning speech, animal vocalizations, and environmental sounds, the study reveals that while frozen Whisper embeddings perform well overall, significant generalization gaps exist for low-resource and non-English languages, with implications for audio AI model development.
