ETRI Journal最新文献

A conventional codec aims to increase the compression efficiency for transmission and storage while maintaining video quality. However, as the number of platforms using machine vision rapidly increases, a codec that increases the compression efficiency and maintains the accuracy of machine vision tasks must be devised. Hence, the Moving Picture Experts Group created a standardization process for video coding for machines (VCM) to reduce bitrates while maintaining the accuracy of machine vision tasks. In particular, in-loop filters have been developed for improving the subjective quality and machine vision task accuracy. However, the high computational complexity of in-loop filters limits the development of a high-performance VCM architecture. We analyze the effect of an in-loop filter on the VCM performance and propose a suboptimal VCM method based on the selective activation of in-loop filters. The proposed method reduces the computation time for video coding by approximately 5% when using the enhanced compression model and 2% when employing a Versatile Video Coding test model while maintaining the machine vision accuracy and compression efficiency of the VCM architecture.

Large language models (LLMs) have revolutionized various applications in natural language processing and exhibited proficiency in generating programming code. We propose a framework for evaluating the code generation ability of LLMs and introduce a new metric, $�p�a�s�s�\text{-}�r�a�t�i�o�@�n$, which captures the granularity of accuracy according to the pass rate of test cases. The framework is intended to be fully automatic to handle the repetitive work involved in generating prompts, conducting inferences, and executing the generated codes. A preliminary evaluation focusing on the prompt detail, problem publication date, and difficulty level demonstrates the successful integration of our framework with the LeetCode coding platform and highlights the applicability of the $�p�a�s�s�\text{-}�r�a�t�i�o�@�n$ metric.

In human conversations, listeners often utilize brief backchannels such as “uh-huh” or “yeah.” Timely backchannels are crucial to understanding and increasing trust among conversational partners. In human–machine conversation systems, users can engage in natural conversations when a conversational agent generates backchannels like a human listener. We propose a method that simultaneously predicts backchannels and recognizes speech in real time. We use a streaming transformer and adopt multitask learning for concurrent backchannel prediction and speech recognition. The experimental results demonstrate the superior performance of our method compared with previous works while maintaining a similar single-task speech recognition performance. Owing to the extremely imbalanced training data distribution, the single-task backchannel prediction model fails to predict any of the backchannel categories, and the proposed multitask approach substantially enhances the backchannel prediction performance. Notably, in the streaming prediction scenario, the performance of backchannel prediction improves by up to 18.7% compared with existing methods.

We introduce a high-performance named entity recognition (NER) model for written and spoken language. To overcome challenges related to labeled data scarcity and domain shifts, we use transfer learning to leverage our previously developed KorBERT as the base model. We also adopt a meta-pseudo-label method using a teacher/student framework with labeled and unlabeled data. Our model presents two modifications. First, the student model is updated with an average loss from both human- and pseudo-labeled data. Second, the influence of noisy pseudo-labeled data is mitigated by considering feedback scores and updating the teacher model only when below a threshold (0.0005). We achieve the target NER performance in the spoken language domain and improve that in the written language domain by proposing a straightforward rollback method that reverts to the best model based on scarce human-labeled data. Further improvement is achieved by adjusting the label vector weights in the named entity dictionary.

Exposure to varied noisy environments impairs the recognition performance of artificial intelligence-based speech recognition technologies. Degraded-performance services can be utilized as limited systems that assure good performance in certain environments, but impair the general quality of speech recognition services. This study introduces an audiovisual speech recognition (AVSR) model robust to various noise settings, mimicking human dialogue recognition elements. The model converts word embeddings and log-Mel spectrograms into feature vectors for audio recognition. A dense spatial–temporal convolutional neural network model extracts features from log-Mel spectrograms, transformed for visual-based recognition. This approach exhibits improved aural and visual recognition capabilities. We assess the signal-to-noise ratio in nine synthesized noise environments, with the proposed model exhibiting lower average error rates. The error rate for the AVSR model using a three-feature multi-fusion method is 1.711%, compared to the general 3.939% rate. This model is applicable in noise-affected environments owing to its enhanced stability and recognition rate.

This paper presents the development of language tutoring systems for non-native speakers by leveraging advanced end-to-end automatic speech recognition (ASR) and proficiency evaluation. Given the frequent errors in non-native speech, high-performance spontaneous speech recognition must be applied. Our systems accurately evaluate pronunciation and speaking fluency and provide feedback on errors by relying on precise transcriptions. End-to-end ASR is implemented and enhanced by using diverse non-native speaker speech data for model training. For performance enhancement, we combine semisupervised and transfer learning techniques using labeled and unlabeled speech data. Automatic proficiency evaluation is performed by a model trained to maximize the statistical correlation between the fluency score manually determined by a human expert and a calculated fluency score. We developed an English tutoring system for Korean elementary students called EBS AI PengTalk and a Korean tutoring system for foreigners called KSI Korean AI Tutor. Both systems were deployed by South Korean government agencies.

We propose a method to automatically predict Alzheimer's disease from speech data using the ChatGPT large language model. Alzheimer's disease patients often exhibit distinctive characteristics when describing images, such as difficulties in recalling words, grammar errors, repetitive language, and incoherent narratives. For prediction, we initially employ a speech recognition system to transcribe participants' speech into text. We then gather opinions by inputting the transcribed text into ChatGPT as well as a prompt designed to solicit fluency evaluations. Subsequently, we extract embeddings from the speech, text, and opinions by the pretrained models. Finally, we use a classifier consisting of transformer blocks and linear layers to identify participants with this type of dementia. Experiments are conducted using the extensively used ADReSSo dataset. The results yield a maximum accuracy of 87.3% when speech, text, and opinions are used in conjunction. This finding suggests the potential of leveraging evaluation feedback from language models to address challenges in Alzheimer's disease recognition.