• ETRI Journal
• /
• v.46 no.1
• /
• pp.48-58
• /
• 2024

This paper presents the development of language tutoring systems for nonnative 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 Peng-Talk and a Korean tutoring system for foreigners called KSI Korean AI Tutor. Both systems were deployed by South Korean government agencies.

• Phonetics and Speech Sciences
• /
• v.15 no.3
• /
• pp.75-82
• /
• 2023

Automatic speech recognition (ASR) has been revolutionized with deep learning-based approaches, among which self-supervised learning methods have proven to be particularly effective. In this study, we aim to enhance the performance of OpenAI's Whisper model, a multilingual ASR system on the Korean language. Whisper was pretrained on a large corpus (around 680,000 hours) of web speech data and has demonstrated strong recognition performance for major languages. However, it faces challenges in recognizing languages such as Korean, which is not major language while training. We address this issue by fine-tuning the Whisper model with an additional dataset comprising about 1,000 hours of Korean speech. We also compare its performance against a Transformer model that was trained from scratch using the same dataset. Our results indicate that fine-tuning the Whisper model significantly improved its Korean speech recognition capabilities in terms of character error rate (CER). Specifically, the performance improved with increasing model size. However, the Whisper model's performance on English deteriorated post fine-tuning, emphasizing the need for further research to develop robust multilingual models. Our study demonstrates the potential of utilizing a fine-tuned Whisper model for Korean ASR applications. Future work will focus on multilingual recognition and optimization for real-time inference.

• Speech Sciences
• /
• v.6
• /
• pp.83-101
• /
• 1999

The aim of this study is to provide a learning method of prosodic rhythm for Taegu North Kyungsang Korean speakers to learn French rhythm more effectively. The rhythmic properties of spoken French and Taegu North Kyungsang Korean dialect are different from each other. Therefore, we try to provide a basic rhythmic model of the two languages by dividing into three parts: syllable, rhythmic unit and accent, and intonation. To do so, we recorded French of Taegu Kyungsang Korean speakers, and then analysed and compared the rhythmic properties of Korean and French by spectrograph. We tried to find rhythmic mistakes in their French pronunciation, and then established a learning model to modify them. After training with the CSL Macro learning model, we observed the output result. However, although learners understand the method we have proposed, an effective method which is possible by repeating practice must be arranged to be actually used in direct verbal communications in a well-developed learning programme. Hence, this study may play an important role at the level of preparation in the setting of an effective rhythmic learning programme.

• Proceedings of the Korean Information Science Society Conference
• /
• 2011.06a
• /
• pp.255-258
• /
• 2011

An application of Gammatone auditory filterbank for emotion recognition from speech is described in this paper. Gammatone filterbank is a bank of Gammatone filters which are used as a preprocessing stage before applying feature extraction methods to get the most relevant features for emotion recognition from speech. In the feature extraction step, the energy value of output signal of each filter is computed and combined with other of all filters to produce a feature vector for the learning step. A feature vector is estimated in a short time period of input speech signal to take the advantage of dependence on time domain. Finally, in the learning step, Hidden Markov Model (HMM) is used to create a model for each emotion class and recognize a particular input emotional speech. In the experiment, feature extraction based on Gammatone filterbank (GTF) shows the better outcomes in comparison with features based on Mel-Frequency Cepstral Coefficient (MFCC) which is a well-known feature extraction for speech recognition as well as emotion recognition from speech.

• Proceedings of the KSPS conference
• /
• 2005.11a
• /
• pp.103-106
• /
• 2005

In this paper, we propose a language modeling approach to improve the performance of a large vocabulary continuous speech recognition system. The proposed approach is based on the active learning framework that helps to select a text corpus from a plenty amount of text data required for language modeling. The perplexity is used as a measure for the corpus selection in the active learning. From the recognition experiments on the task of continuous Korean speech, the speech recognition system employing the language model by the proposed language modeling approach reduces the word error rate by about 6.6 % with less computational complexity than that using a language model constructed with randomly selected texts.

• ETRI Journal
• /
• v.46 no.1
• /
• pp.118-126
• /
• 2024

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.

• Phonetics and Speech Sciences
• /
• v.15 no.3
• /
• pp.69-74
• /
• 2023

This study introduces an innovative model for zero-shot voice conversion that utilizes the capabilities of HuBERT. Zero-shot voice conversion models can transform the speech of one speaker to mimic that of another, even when the model has not been exposed to the target speaker's voice during the training phase. Comprising five main components (HuBERT, feature encoder, flow, speaker encoder, and vocoder), the model offers remarkable performance across a range of scenarios. Notably, it excels in the challenging unseen-to-unseen voice-conversion tasks. The effectiveness of the model was assessed based on the mean opinion scores and similarity scores, reflecting high voice quality and similarity to the target speakers. This model demonstrates considerable promise for a range of real-world applications demanding high-quality voice conversion. This study sets a precedent in the exploration of HuBERT-based models for voice conversion, and presents new directions for future research in this domain. Despite its complexities, the robust performance of this model underscores the viability of HuBERT in advancing voice conversion technology, making it a significant contributor to the field.

• Phonetics and Speech Sciences
• /
• v.8 no.2
• /
• pp.17-22
• /
• 2016

This paper devises a model to diagnose physical fatigue using speech features. This paper presents a machine learning method through an SVM algorithm using the various feature parameters. The parameters used include the significant speech parameters, questionnaire responses, and bio-signal parameters obtained before and after the experiment imposing the fatigue. The results showed that performance rates of 95%, 100%, and 90%, respectively, were observed from the proposed model using three types of the parameters relevant to the fatigue. These results suggest that the method proposed in this study can be used as the physical fatigue diagnostic model, and that fatigue can be easily diagnosed by speech technology.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.22 no.3
• /
• pp.69-74
• /
• 2022

SNN (Spiking Neural Network) operating in neuromorphic architecture was created by mimicking human neural networks. Neuromorphic computing based on neuromorphic architecture requires relatively lower power than typical deep learning techniques based on GPUs. For this reason, research to support various artificial intelligence models using neuromorphic architecture is actively taking place. This paper conducted a performance analysis of the speech recognition model based on neuromorphic architecture according to the speech data preprocessing technique. As a result of the experiment, it showed up to 84% of speech recognition accuracy performance when preprocessing speech data using the Fourier transform. Therefore, it was confirmed that the speech recognition service based on the neuromorphic architecture can be effectively utilized.

• The Journal of Bigdata
• /
• v.8 no.2
• /
• pp.223-231
• /
• 2023

In this study, we propose a model that extracts and analyzes features from deep learning-based speech signals, generates filters, and utilizes these filters to recognize emotions in speech signals. We evaluate the performance of emotion recognition accuracy using the proposed model. According to the simulation results using the proposed model, the average emotion recognition accuracy of DNN and RNN was very similar, at 84.59% and 84.52%, respectively. However, we observed that the simulation time for DNN was approximately 44.5% shorter than that of RNN, enabling quicker emotion prediction.