• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.20 no.3
• /
• pp.464-470
• /
• 2016

This paper presents a speech recognition database verification system using speech measures, and describes a speech measure extraction algorithm which is applied to this system. In our previous study, to produce an effective speech quality measure for the system, we propose a combination of various speech measures which are highly correlated with WER (Word Error Rate). The new combination of various types of speech quality measures in this study is more effective to predict the speech recognition performance compared to each speech measure alone. In this paper, we increase the system independency by employing GMM acoustic score instead of HMM score which is obtained by a secondary speech recognition system. The combination with GMM score shows a slightly lower correlation with WER compared to the combination with HMM score, however it presents a higher relative improvement in correlation with WER, which is calculated compared to the correlation of each speech measure alone.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.20 no.3
• /
• pp.471-476
• /
• 2016

This paper presents our study on speech recognition performance prediction. Our initial study shows that a combination of speech quality measures effectively improves correlation with Word Error Rate (WER) compared to each speech measure alone. In this paper we demonstrate a new combination of various types of speech quality measures shows more significantly improves correlation with WER compared to the speech measure combination of our initial study. In our study, SNR, PESQ, acoustic model score, and MFCC distance are used as the speech quality measures. This paper also presents our speech database verification system for speech recognition employing the speech measures. We develop a WER prediction system using Gaussian mixture model and the speech quality measures as a feature vector. The experimental results show the proposed system is highly effective at predicting WER in a low SNR condition of speech babble and car noise environments.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.21 no.12
• /
• pp.2291-2297
• /
• 2017

This paper describe to extract speech measure algorithm for evaluating a speech database, and presents generating method of a speech quality measure using DNN(Deep Neural Network). In our previous study, to produce an effective speech quality measure, we propose a combination of various speech measures which are highly correlated with WER(Word Error Rate). The new combination of various types of speech quality measures in this study is more effective to predict the speech recognition performance compared to each speech measure alone. In this paper, we describe the method of extracting measure using DNN, and we change one of the combined measure from GMM(Gaussican Mixture Model) score used in the previous study to DNN score. The combination with DNN score shows a higher correlation with WER compared to the combination with GMM score.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.52-54
• /
• 2023

최근 인공지능 기술의 발전에 따라서 다양한 서비스에서 음성 인식을 활용한 서비스를 제공하면서 음성 인식에 대한 중요성이 증가하고 있다. 이 논문에서는 국내에서 많이 사용되고 있는 대표적인 인공지능 서비스 API 를 제공하는 구글, ETRI, 네이버에 대해서 감정 음성 관점에서 그 차이를 평가하였다. AI Hub 에서 제공하는 감성 대화 말뭉치 데이터 셋의 일부인 음성 테스트 데이터를 사용하여 평가한 결과 ETRI API 가 문자 오류율 (1.29%)과 단어 오류율(10.1%)의 성능 지표에 대해서 가장 우수한 음성 인식 성능을 보임을 확인하였다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.716-721
• /
• 2023

Inverse Text Normalization(ITN) 모델은 음성 인식(STT) 엔진의 중요한 후처리 영역 중 하나이며, STT 인식 결과의 가독성을 개선한다. 최근 ITN 모델에 심층신경망을 활용한 연구가 진행되고 있다. 심층 신경망을 사용하는 대부분의 선행연구는 문장 내 변환이 필요한 부분에 토큰 태깅을 진행하는 방식이다. 그러나 이는 Out-of-vocabulary(OOV) 이슈가 있으며, 학습 데이터 구축 시 토큰 단위의 섬세한 태깅 작업이 필요하다는 한계점이 존재한다. 더불어 선행 연구에서는 STT 인식 결과를 그대로 사용하는데, 이는 띄어쓰기가 중요한 한국어 ITN 처리에 변환 성능을 보장할 수 없다. 본 연구에서는 BART 기반 생성 모델로 생성형 ITN 모델을 구축하였고, Bi-LSTM 기반 보조 신경망 모델을 결합하여 STT 인식 결과에 대한 고유명사 처리, 띄어쓰기 교정 기능을 보완한 모델을 제안한다. 또한 보조 신경망을 통해 생성 모델 처리 여부를 판단하여 평균 추론 속도를 개선하였다. 실험을 통해 두 모델의 각 정량 성능 지표에서 우수한 성능을 확인하였고 결과적으로 본 연구에서 제안하는 두 모델의 결합된 방법론의 효과성을 제시하였다.

• The Journal of the Convergence on Culture Technology
• /
• v.10 no.1
• /
• pp.617-623
• /
• 2024

Identifying emotions through speech poses a significant challenge due to the complex relationship between language and emotions. Our paper aims to take on this challenge by employing feature engineering to identify emotions in speech through a multimodal classification task involving both speech and text data. We evaluated two classifiers-Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM)-both integrated with a BERT-based pre-trained model. Our assessment covers various performance metrics (accuracy, F-score, precision, and recall) across different experimental setups). The findings highlight the impressive proficiency of two models in accurately discerning emotions from both text and speech data.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.17 no.7
• /
• pp.945-951
• /
• 2007

Noise cancellation is essential for the devices which use speech as an interface. In real environments, speech quality and recognition rates are degraded by the auditive noises coming near the microphone. In this paper, we propose a noise cancellation algorithm using stereo microphones basically. The advantage of the use of multiple microphones is that the direction information of the target source could be applied. The proposed noise canceller is based on the Wiener filter. To estimate the filter, noise and target speech frequency responses should be known and they are estimated by the spectral classification in the frequency domain. The performance of the proposed algorithm is compared with that of the well-known Frost algorithm and the generalized sidelobe canceller (GSC) with an adaptation mode controller (AMC). As performance measures, the perceptual evaluation of speech quality (PESQ), which is the most widely used among various objective speech quality methods, and speech recognition rates are adopted.

• Smart Media Journal
• /
• v.12 no.5
• /
• pp.28-35
• /
• 2023

In this study, we propose a new light-weight model RoutingConvNet with fewer parameters to improve the applicability and practicality of speech emotion recognition. To reduce the number of learnable parameters, the proposed model connects bidirectional MFCCs on a channel-by-channel basis to learn long-term emotion dependence and extract contextual features. A light-weight deep CNN is constructed for low-level feature extraction, and self-attention is used to obtain information about channel and spatial signals in speech signals. In addition, we apply dynamic routing to improve the accuracy and construct a model that is robust to feature variations. The proposed model shows parameter reduction and accuracy improvement in the overall experiments of speech emotion datasets (EMO-DB, RAVDESS, and IEMOCAP), achieving 87.86%, 83.44%, and 66.06% accuracy respectively with about 156,000 parameters. In this study, we proposed a metric to calculate the trade-off between the number of parameters and accuracy for performance evaluation against light-weight.

• Journal of Digital Contents Society
• /
• v.17 no.2
• /
• pp.71-79
• /
• 2016

This study is intended to compare the performances of sound source localization methods used for stable automobile control by improving voice recognition rate in automobile environment and suggest how to improve their performances. Generally, sound source location estimation methods employ the TDOA algorithm, and there are two ways for it; one is to use a cross correlation function in the time domain, and the other is GCC-PHAT calculated in the frequency domain. Among these ways, GCC-PHAT is known to have stronger characteristics against echo and noise than the cross correlation function. This study compared the performances of the two methods above in automobile environment full of echo and vibration noise and suggested the use of a median filter additionally. We found that median filter helps both estimation methods have good performances and variance values to be decreased. According to the experimental results, there is almost no difference in the two methods' performances in the experiment using voice; however, using the signal of a song, GCC-PHAT is 10% more excellent than the cross correlation function in terms of the recognition rate. Also, when the median filter was added, the cross correlation function's recognition rate could be improved up to 11%. And in regarding to variance values, both methods showed stable performances.

• Journal of Digital Convergence
• /
• v.19 no.3
• /
• pp.203-209
• /
• 2021

In this study, a prototype design model was proposed for developing an online recruitment system through multi-dimensional data crawling and social media analysis, and validates text information and video interview in job application process. This study includes a comparative analysis process through text mining to verify the authenticity of job application paperwork and to effectively hire and allocate workers based on the potential job capability. Based on the prototype system, we conducted performance tests and analyzed the result for key performance indicators such as text mining accuracy and interview STT(speech to text) function recognition rate. If commercialized based on design specifications and prototype development results derived from this study, it may be expected to be utilized as the intelligent online recruitment system technology required in the public and private recruitment markets in the future.