• Phonetics and Speech Sciences
• /
• v.12 no.2
• /
• pp.29-37
• /
• 2020

Recently, the neural network-based deep learning algorithm has dramatically improved performance compared to the classical Gaussian mixture model based hidden Markov model (GMM-HMM) automatic speech recognition (ASR) system. In addition, researches on end-to-end (E2E) speech recognition systems integrating language modeling and decoding processes have been actively conducted to better utilize the advantages of deep learning techniques. In general, E2E ASR systems consist of multiple layers of encoder-decoder structure with attention. Therefore, E2E ASR systems require data with a large amount of speech-text paired data in order to achieve good performance. Obtaining speech-text paired data requires a lot of human labor and time, and is a high barrier to building E2E ASR system. Therefore, there are previous studies that improve the performance of E2E ASR system using relatively small amount of speech-text paired data, but most studies have been conducted by using only speech-only data or text-only data. In this study, we proposed a semi-supervised training method that enables E2E ASR system to perform well in corpus in different domains by using both speech or text only data. The proposed method works effectively by adapting to different domains, showing good performance in the target domain and not degrading much in the source domain.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.6
• /
• pp.578-586
• /
• 2021

We propose a semi-supervised learning method based on Variational AutoEncoder (VAE) and Unsupervised Data Augmentation (UDA) to improve the performance of an end-to-end speech recognizer. In the proposed method, first, the VAE-based augmentation model and the baseline end-to-end speech recognizer are trained using the original speech data. Then, the baseline end-to-end speech recognizer is trained again using data augmented from the learned augmentation model. Finally, the learned augmentation model and end-to-end speech recognizer are re-learned using the UDA-based semi-supervised learning method. As a result of the computer simulation, the augmentation model is shown to improve the Word Error Rate (WER) of the baseline end-to-end speech recognizer, and further improve its performance by combining it with the UDA-based learning method.

• Phonetics and Speech Sciences
• /
• v.11 no.1
• /
• pp.41-49
• /
• 2019

An end-to-end speech recognition model consisting of a single integrated neural network model was recently proposed. The end-to-end model does not need several training steps, and its structure is easy to understand. However, it is difficult to understand how the model recognizes speech internally. In this paper, we visualized and analyzed the attention-based end-to-end model to elucidate its internal mechanisms. We compared the acoustic model of the BLSTM-HMM hybrid model with the encoder of the end-to-end model, and visualized them using t-SNE to examine the difference between neural network layers. As a result, we were able to delineate the difference between the acoustic model and the end-to-end model encoder. Additionally, we analyzed the decoder of the end-to-end model from a language model perspective. Finally, we found that improving end-to-end model decoder is necessary to yield higher performance.

• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.145-150
• /
• 2020

최근 음성인식 분야에서 신경망 기반의 종단간 모델이 제안되고 있다. 해당 모델들은 음성을 직접 입력받아 전사된 문장을 생성한다. 음성을 직접 입력받는 모델의 특성상 데이터의 품질이 모델의 성능에 많은 영향을 준다. 본 논문에서는 이러한 종단간 모델의 문제점을 해결하고자 음성인식 결과를 후처리하기 위한 멀티모달 기반 모델을 제안한다. 제안 모델은 음성과 전사된 문장을 입력 받는다. 입력된 각각의 데이터는 Encoder를 통해 자질을 추출하고 주의집중 메커니즘을 통해 Decoder로 추출된 정보를 전달한다. Decoder에서는 전달받은 주의집중 메커니즘의 결과를 바탕으로 후처리된 토큰을 생성한다. 본 논문에서는 후처리 모델의 성능을 평가하기 위해 word error rate를 사용했으며, 실험결과 Google cloud speech to text모델에 비해 word error rate가 8% 감소한 것을 확인했다.

• Phonetics and Speech Sciences
• /
• v.12 no.4
• /
• pp.63-71
• /
• 2020

Speech recognition is one of the areas actively commercialized using deep learning and machine learning techniques. However, the majority of speech recognition systems on the market are developed on data with limited diversity of speakers and tend to perform well on typical adult speakers only. This is because most of the speech recognition models are generally learned using a speech database obtained from adult males and females. This tends to cause problems in recognizing the speech of the elderly, children and people with dialects well. To solve these problems, it may be necessary to retain big database or to collect a data for applying a speaker adaptation. However, this paper proposes that a new end-to-end speech recognition method consists of an acoustic augmented recurrent encoder and a transformer decoder with linguistic prediction. The proposed method can bring about the reliable performance of acoustic and language models in limited data conditions. The proposed method was evaluated to recognize Korean elderly and children speech with limited amount of training data and showed the better performance compared of a conventional method.

• The Korean Journal of Applied Statistics
• /
• v.32 no.2
• /
• pp.213-227
• /
• 2019

In this paper, we propose an end-to-end deep learning model combining Bayesian neural network with Korean speech recognition. In the past, Korean speech recognition was a complicated task due to the excessive parameters of many intermediate steps and needs for Korean expertise knowledge. Fortunately, Korean speech recognition becomes manageable with the aid of recent breakthroughs in "End-to-end" model. The end-to-end model decodes mel-frequency cepstral coefficients directly as text without any intermediate processes. Especially, Connectionist Temporal Classification loss and Attention based model are a kind of the end-to-end. In addition, we combine Bayesian neural network to implement the end-to-end model and obtain Monte Carlo estimates. Finally, we carry out our experiments on the "WorimalSam" online dictionary dataset. We obtain 4.58% Word Error Rate showing improved results compared to Google and Naver API.

• Electronics and Telecommunications Trends
• /
• v.19 no.3 s.87
• /
• pp.136-144
• /
• 2004

인터넷의 발달 및 VoIP의 보급으로 인해 VoIP 서비스의 품질에 대한 관심이 증가하고 있다. 그 동안은 망사업자 관점에서 망의 품질을 개선하기 위한 MPLS, Diffserv, RSVP 등의 연구가 진행되어 왔으나, 실제로 서비스 품질은 망뿐만 아니라 단말 등의 품질에도 영향을 받기 때문에 망 사업자의 관점에서 보는 서비스 품질 기준이 아닌, 고객의 관점에서 인식 가능한 수준에서의 종단간 서비스 품질을 다룰 필요가 있다. 본 고에서는 서비스 품질이란 무엇인지 살펴보고, 국제표준단체의 서비스 품질 관련 연구 및 VoIP 서비스를 위한 음성 품질 평가 기술에 대하여 살펴본다.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.22 no.12
• /
• pp.1604-1610
• /
• 2018

The speech recognition service is used as an ancillary means of communication by converting and visualizing the speaker's voice into text to the hearing impaired persons. However, in open environments such as classrooms and conference rooms it is difficult to provide speech recognition service to many hearing impaired persons. For this, a method is needed to efficiently provide it according to the surrounding environment. In this paper, we propose a system that recognizes the speaker's voice and transmits the converted text to many hearing impaired persons as messages. The proposed system uses the MQTT protocol to deliver messages to many users at the same time. The end-to-end delay was measured to confirm the service delay of the proposed system according to the QoS level setting of the MQTT protocol. As a result of the measurement, the delay between the most reliable Qos level 2 and 0 is 111ms, confirming that it does not have a great influence on conversation recognition.

• The Journal of the Acoustical Society of Korea
• /
• v.41 no.2
• /
• pp.122-129
• /
• 2022

In this paper, we propose a method to improve the performance of Transformer-based speech recognizers using data augmentation that locally adjusts the frame rate. First, the start time and length of the part to be augmented in the original voice data are randomly selected. Then, the frame rate of the selected part is changed to a new frame rate by using linear interpolation. Experimental results using the Wall Street Journal and LibriSpeech speech databases showed that the convergence time took longer than the baseline, but the recognition accuracy was improved in most cases. In order to further improve the performance, various parameters such as the length and the speed of the selected parts were optimized. The proposed method was shown to achieve relative performance improvement of 11.8 % and 14.9 % compared with the baseline in the Wall Street Journal and LibriSpeech speech databases, respectively.