• 대한음성학회:학술대회논문집
• /
• 대한음성학회 2005년도 춘계 학술대회 발표논문집
• /
• pp.99-102
• /
• 2005

Research efforts have been made for out-of-vocabulary word rejection to improve the confidence of speech recognition systems. However, little attention has been paid to non-recognition sentence rejection. According to the appearance of pronunciation correction systems using speech recognition technology, it is needed to reject non-recognition sentences to provide users with more accurate and robust results. In this paper, we introduce standard phoneme based sentence rejection system with no need of special filler models. Instead we used word spotting ratio to determine whether input sentences would be accepted or rejected. Experimental results show that we can achieve comparable performance using only standard phoneme based recognition network in terms of the average of FRR and FAR.

• 대한음성학회:학술대회논문집
• /
• 대한음성학회 2003년도 5월 학술대회지
• /
• pp.219-222
• /
• 2003

In this paper, we study the usage of phoneme duration information for rejection garbage sentence. First, we build a phoneme duration modeling in a speech recognition system based on dicicion tree state tying, We assume that phone duration has a Gamma distribution. Next, we build a verification module in which word-level confidence measure is used. Finally, we make a comparative study on phoneme duration with speech DB obtained from the live system. This DB consistes of OOT(out-of-task) and ING(in-grammar) utterences. the usage of phone duration information yields that OOT recognition rate is improved by 46％ and that another 8.4％ error rate is reduced when combined with utterence verification module.

• ETRI Journal
• /
• 제28권6호
• /
• pp.811-814
• /
• 2006

In this letter, we introduce confusion-based confidence measures for detecting an impostor in speaker recognition, which does not require an alternative hypothesis. Most traditional speaker verification methods are based on a hypothesis test, and their performance depends on the robustness of an alternative hypothesis. Compared with the conventional Gaussian mixture model-universal background model (GMM-UBM) scheme, our confusion-based measures show better performance in noise-corrupted speech. The additional computational requirements for our methods are negligible when used to detect or reject impostors.

• 대한음성학회:학술대회논문집
• /
• 대한음성학회 2005년도 추계 학술대회 발표논문집
• /
• pp.215-218
• /
• 2005

This paper reports the current status of development of the Korean speech recognition platform (ECHOS). We implement new modules including ETSI feature extraction, backward search with trigram, and utterance verification. The ETSI feature extraction module is implemented by converting the public software to an object-oriented program. We show that trigram language modeling in the backward search pass reduces the word error rate from 23.5% to 22% on a large vocabulary continuous speech recognition task. We confirm the utterance verification module by examining word graphs with confidence score.

• 음성과학
• /
• 제12권3호
• /
• pp.105-114
• /
• 2005

It was verified that a speaker verification system improved its performances of EER by regularizing log likelihood ratio, using background speaker models. Recently the wireless mobile phones are becoming more dominant communication terminals than wired phones. So the need for building a speaker verification system on mobile phone is increasing abruptly. Therefore in this paper, we had some experiments to examine the performance of speaker verification based on mobile phone's voices. Especially we are focused on the performance variations in EER(Equal Error Rate) according to several background speaker's characteristics, such as selecting methods(MSC, MIX), number of background speakers, aging factor of speech database. For this, we constructed a speaker verification system that uses GMM(Gaussin Mixture Model) and found that the MIX method is generally superior to another method by about 1.0% EER. In aspect of number of background speakers, EER is decreasing in proportion to the background speakers populations. As the number is increasing as 6, 10 and 16, the EERs are recorded as 13.0%, 12.2%, and 11.6%. An unexpected results are happened in aging effects of the speech database on the performance. EERs are measured as 4%, 12% and 19% for each seasonally recorded databases from session 1 to session 3, respectively, where duration gap between sessions is set by 3 months. Although seasons speech database has 10 speakers and 10 sentences per each, which gives less statistical confidence to results, we confirmed that enrolled speaker models in speaker verification system should be regularly updated using the ongoing claimant's utterances.

• 말소리와 음성과학
• /
• 제5권2호
• /
• pp.43-51
• /
• 2013

English native listeners have a tendency to treat strong syllables in a speech stream as the potential initial syllables of new words, since the majority of lexical words in English have a word-initial stress. The current study investigates whether Korean (L1) - English (L2) late bilinguals perceive strong syllables in English continuous speech as word onsets, as English native listeners do. In Experiment 1, word-spotting was slower when the word-initial syllable was strong, indicating that Korean listeners do not perceive strong syllables as word onsets. Experiment 2 was conducted in order to avoid any possibilities that the results of Experiment 1 may be due to the strong-initial targets themselves used in Experiment 1 being slower to recognize than the weak-initial targets. We employed the gating paradigm in Experiment 2, and measured the Isolation Point (IP, the point at which participants correctly identify a word without subsequently changing their minds) and the Recognition Point (RP, the point at which participants correctly identify the target with 85% or greater confidence) for the targets excised from the non-words in the two conditions of Experiment 1. Both the mean IPs and the mean RPs were significantly earlier for the strong-initial targets, which means that the results of Experiment 1 reflect the difficulty of segmentation when the initial syllable of words was strong. These results are consistent with Kim & Nam (2011), indicating that strong syllables are not perceived as word onsets for Korean listeners and interfere with lexical segmentation in English running speech.

• 한국음향학회지
• /
• 제20권2호
• /
• pp.27-34
• /
• 2001

발화 검증이란 등록된 단어 목록 이외의 단어가 입력되었을 때, 미등록된 단어는 인식할 수 없는 단어임을 알려주는 기능으로써 사용자에게 친숙한 음성 인식 시스템을 설계하는데 중요한 기술이다. 본 논문에서는 가변어휘 단어 인식기에서 최소 검증 오류를 나타낼 수 있는 발화 검증 시스템의 알고리즘을 제안한다. 우선, 한국전자통신연구원의 PBW(Phonetically Balanced Words) 445DB를 이용하여 가변어휘 단어 인식에서의 미등록어 거절 성능을 향상시키는 효과적인 발화 검증 방법을 제안하였다. 구체적으로 특별한 훈련 과정이 없이도 유사 음소 집합을 많이 포함시킨 반음소 모델을 제안하여 최소 검증 오류를 지니도록 하였다. 또한, 음소 단위의 null hypothesis와 alternate hypothesis의 비를 이용한 음소 단위의 신뢰도는 null hypothesis로 정규화해서 강인한 발화 검증 성능을 보여 주었으며, 음소 단위의 신뢰도를 이용한 단어 단위의 신뢰도는 등록어와 미등록어 사이의 분별력을 잘 표현해 주었다. 이와 같이 새로이 제안된 반음소 모델과 발화 검증 방법을 사용했을 때, CA (Correctly Accept for Keyword: 등록어를 제대로 인정한 경우)는 약 89％, CR (Correctly Reject for OOV (Out-of-Vocabulary): 미등록어에 대해 거절한 경우)은 약 90％로써, 기존 필터 모델을 이용한 방법보다 미등록어 거절 성능이 ERR (Error Reduction Rate) 측면에서 약 15-21％ 향상됨을 알 수 있었다.

• 음성과학
• /
• 제14권1호
• /
• pp.163-174
• /
• 2007

All observation has uncertainty due to noise or channel characteristics. This uncertainty should be counted in the modeling of observation. In this paper we propose a modified optimization object function of a GMM training considering inexact observation. The object function is modified by introducing the concept of observation confidence as a weighting factor of probabilities. The optimization of the proposed criterion is solved using a common EM algorithm. To verify the proposed method we apply it to the speaker recognition domain. The experimental results of text-independent speaker identification with VidTimit DB show that the error rate is reduced from 14.8% to 11.7% by the modified GMM training.

• 대한음성학회지:말소리
• /
• 제56호
• /
• pp.147-158
• /
• 2005

In this paper, we are focused to develop the simplified speaker verification algorithm without background speaker models, which will be adopted in the portable speaker verification system equipped in portable terminals such as mobile phone and PMP. According to the tolerance interval analysis, the population of someone's speaker model can be represented by a suitable number of selected independent samples of speaker model. So we can make the representative speaker model and threshold under the specified confidence level and coverage. Using proposed algorithm with the number of samples is 40, the experiments show that the false rejection rate is $3.0\%$ and the false acceptance rate $4.3\%$, worth comparing to conventional method's results, $5.4\%\;and\;5.5\%$, respectively. Next step of research will be on the suitable adaptation methods to overcome speech variation problems due to aging effect and operating environments.

• 대한음성학회지:말소리
• /
• 제66호
• /
• pp.87-103
• /
• 2008

We describe a new active learning method in conditional random fields (CRFs) framework for keyphrase extraction. To save elaboration in annotation, we use diversity and representative measure. We select high diversity training candidates by sentence confidence value. We also select high representative candidates by clustering the part-of-speech patterns of contexts. In the experiments using dialog corpus, our method achieves 86.80% and saves 88% training corpus compared with those of supervised method. From the results of experiment, we can see that the proposed method shows improved performance over the previous methods. Additionally, the proposed method can be applied to other applications easily since its implementation is independent on applications.