• Journal of Digital Contents Society
• /
• v.18 no.1
• /
• pp.93-99
• /
• 2017

A hybrid approach using Long Short Term Memory (LSTM) Recurrent Neural Network (RNN) has showed great improvement in speech recognition accuracy. For training acoustic model based on hybrid approach, it requires forced alignment of HMM state sequence from Gaussian Mixture Model (GMM)-Hidden Markov Model (HMM). However, high computation time for training GMM-HMM is required. This paper proposes an end-to-end approach for LSTM RNN-based Korean speech recognition to improve learning speed. A Connectionist Temporal Classification (CTC) algorithm is proposed to implement this approach. The proposed method showed almost equal performance in recognition rate, while the learning speed is 1.27 times faster.

• ETRI Journal
• /
• v.44 no.3
• /
• pp.476-490
• /
• 2022

With the recent advances in technology, automatic speech recognition (ASR) has been widely used in real-world applications. The efficiency of converting large amounts of speech into text accurately with limited resources has become more vital than ever. In this study, we propose a method to rapidly recognize a large speech database via a transformer-based end-to-end model. Transformers have improved the state-of-the-art performance in many fields. However, they are not easy to use for long sequences. In this study, various techniques to accelerate the recognition of real-world speeches are proposed and tested, including decoding via multiple-utterance-batched beam search, detecting end of speech based on a connectionist temporal classification (CTC), restricting the CTC-prefix score, and splitting long speeches into short segments. Experiments are conducted with the Librispeech dataset and the real-world Korean ASR tasks to verify the proposed methods. From the experiments, the proposed system can convert 8 h of speeches spoken at real-world meetings into text in less than 3 min with a 10.73% character error rate, which is 27.1% relatively lower than that of conventional systems.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.414-418
• /
• 2021

다국어 음성인식은 단일언어 음성인식에 비해 높은 난이도를 보인다. 하나의 단일 모델로 다국어 음성인식을 수행하기 위해선 다양한 언어가 공유하는 음성적 특성을 모델이 학습할 수 있도록 하여 음성인식 성능을 향상시킬 수 있다. 본 연구는 딥러닝 음성인식 모델인 Wav2Vec2.0 구조를 변경하여 한국어와 영어 음성을 하나의 모델로 학습하는 방법을 제시한다. CTC(Connectionist Temporal Classification) 손실함수를 이용하는 Wav2Vec2.0 모델의 구조에서 각 언어마다 별도의 CTC 출력 계층을 두고 각 언어별 사전(Lexicon)을 적용하여 음성 입력을 다른 언어로 혼동되는 경우를 원천적으로 방지한다. 제시한 Wav2Vec2.0 구조를 사용하여 한국어와 영어를 잘못 분류하여 음성인식률이 낮아지는 문제를 해결하고 더불어 제시된 한국어 음성 데이터셋(KsponSpeech)에서 한국어와 영어를 동시에 학습한 모델이 한국어만을 이용한 모델보다 향상된 음성 인식률을 보임을 확인하였다. 마지막으로 Prefix 디코딩을 활용하여 언어모델을 이용한 음성인식 성능 개선을 수행하였다.

• 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.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.18 no.6
• /
• pp.1692-1705
• /
• 2024

In this paper, we present a method that integrates a Grammar Transducer as an external language model to enhance the accuracy of the pre-trained Korean End-to-end (E2E) Automatic Speech Recognition (ASR) model. The E2E ASR model utilizes the Connectionist Temporal Classification (CTC) loss function to derive hypothesis sentences from input audio. However, this method reveals a limitation inherent in the CTC approach, as it fails to capture language information from transcript data directly. To overcome this limitation, we propose a fusion approach that combines a clause-level n-gram language model, transformed into a Weighted Finite-State Transducer (WFST), with the E2E ASR model. This approach enhances the model's accuracy and allows for domain adaptation using just additional text data, avoiding the need for further intensive training of the extensive pre-trained ASR model. This is particularly advantageous for Korean, characterized as a low-resource language, which confronts a significant challenge due to limited resources of speech data and available ASR models. Initially, we validate the efficacy of training the n-gram model at the clause-level by contrasting its inference accuracy with that of the E2E ASR model when merged with language models trained on smaller lexical units. We then demonstrate that our approach achieves enhanced domain adaptation accuracy compared to Shallow Fusion, a previously devised method for merging an external language model with an E2E ASR model without necessitating additional training.

• KIPS Transactions on Software and Data Engineering
• /
• v.8 no.3
• /
• pp.115-122
• /
• 2019

For Korean phoneme recognition, Hidden Markov-Gaussian Mixture model(HMM-GMM) or hybrid models which combine artificial neural network with HMM have been mainly used. However, current approach has limitations in that such models require force-aligned corpus training data that is manually annotated by experts. Recently, researchers used neural network based phoneme recognition model which combines recurrent neural network(RNN)-based structure with connectionist temporal classification(CTC) algorithm to overcome the problem of obtaining manually annotated training data. Yet, in terms of implementation, these RNN-based models have another difficulty in that the amount of data gets larger as the structure gets more sophisticated. This problem of large data size is particularly problematic in the Korean language, which lacks refined corpora. In this study, we introduce CTC algorithm that does not require force-alignment to create a Korean phoneme recognition model. Specifically, the phoneme recognition model is based on convolutional neural network(CNN) which requires relatively small amount of data and can be trained faster when compared to RNN based models. We present the results from two different experiments and a resulting best performing phoneme recognition model which distinguishes 49 Korean phonemes. The best performing phoneme recognition model combines CNN with 3hop Bidirectional LSTM with the final Phoneme Error Rate(PER) at 3.26. The PER is a considerable improvement compared to existing Korean phoneme recognition models that report PER ranging from 10 to 12.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2018.05a
• /
• pp.398-401
• /
• 2018

본 연구에서는 심충 합성곱 신경망(Deep CNN)과 Connectionist Temporal Classification (CTC) 알고리즘을 사용하여 강제정렬 (force-alignment)이 이루어진 코퍼스 없이도 학습이 가능한 음소 인식 모델을 제안한다. 최근 해외에서는 순환 신경망(RNN)과 CTC 알고리즘을 사용한 딥 러닝 기반의 음소 인식 모델이 활발히 연구되고 있다. 하지만 한국어 음소 인식에는 HMM-GMM 이나 인공 신경망과 HMM 을 결합한 하이브리드 시스템이 주로 사용되어 왔으며, 이 방법 은 최근의 해외 연구 사례들보다 성능 개선의 여지가 적고 전문가가 제작한 강제정렬 코퍼스 없이는 학습이 불가능하다는 단점이 있다. 또한 RNN 은 학습 데이터가 많이 필요하고 학습이 까다롭다는 단점이 있어, 코퍼스가 부족하고 기반 연구가 활발하게 이루어지지 않은 한국어의 경우 사용에 제약이 있다. 이에 본 연구에서는 강제정렬 코퍼스를 필요로 하지 않는 CTC 알고리즘을 도입함과 동시에, RNN 에 비해 더 학습 속도가 빠르고 더 적은 데이터로도 학습이 가능한 합성곱 신경망(CNN)을 사용하여 딥 러닝 모델을 구축하여 한국어 음소 인식을 수행하여 보고자 하였다. 이 모델을 통해 본 연구에서는 한국어에 존재하는 49 가지의 음소를 추출하는 세 종류의 음소 인식기를 제작하였으며, 최종적으로 선정된 음소 인식 모델의 PER(phoneme Error Rate)은 9.44 로 나타났다. 선행 연구 사례와 간접적으로 비교하였을 때, 이 결과는 제안하는 모델이 기존 연구 사례와 대등하거나 조금 더 나은 성능을 보인다고 할 수 있다.

• Phonetics and Speech Sciences
• /
• v.13 no.1
• /
• pp.45-51
• /
• 2021

End-to-end (E2E) automatic speech recognition (ASR) has achieved promising performance gains with the introduced self-attention network, Transformer. However, due to training time and the number of hyperparameters, finding the optimal hyperparameter set is computationally expensive. This paper investigates the impact of hyperparameters in the Transformer network to answer two questions: which hyperparameter plays a critical role in the task performance and training speed. The Transformer network for training has two encoder and decoder networks combined with Connectionist Temporal Classification (CTC). We have trained the model with Wall Street Journal (WSJ) SI-284 and tested on devl93 and eval92. Seventeen hyperparameters were selected from the ESPnet training configuration, and varying ranges of values were used for experiments. The result shows that "num blocks" and "linear units" hyperparameters in the encoder and decoder networks reduce Word Error Rate (WER) significantly. However, performance gain is more prominent when they are altered in the encoder network. Training duration also linearly increased as "num blocks" and "linear units" hyperparameters' values grow. Based on the experimental results, we collected the optimal values from each hyperparameter and reduced the WER up to 2.9/1.9 from dev93 and eval93 respectively.