• Proceedings of the Acoustical Society of Korea Conference
• /
• 1996.10a
• /
• pp.56-60
• /
• 1996

본 논문에서 핵심어 검출(Keyword spotting ) 시스템에서의 채널 왜곡에 대한 보상방법등의 성능을 비교하였다. 훈련을 음성과 인식실험용 음성은 서로 다른 환경에서 수집되었으며, 특별히 인식실험용 음성으로는 전화망을 통한 음성 데이터를 이용하였다. 전화망을 통한 음성인식에서는 채널왜곡과 부가잡음에 의해서 음성신호에 왜곡이 생기므로 이들에 대한 적적한 보상이 필요하다. 본 논문에서는 채널 왜곡보상을 위한 처리방법으로 널리 사용되고 있는 global cepstral mean substraction (GCMS), local cepstral mean subtraction(LCMS) 그리고 RASTA processing을 적용하였다. 그리고 인식성능의 개선을 위해 이들 방법을 likelihood ration scorning 에 의한 후처리 과정을 적용하였다. 인식실험결과 이들 방법 모두 채널왜곡 보상을 하지 않았을 경우와 비교하여 더 좋은 인식성능을 얻을 수 있었으며, 그 중 후처리를 적용한 LCMS 방법이 가장 우수한 성능을 나타내었다.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.4
• /
• pp.380-386
• /
• 2002

Conventional post-processing as like confidence measure (CM) proposed by Rahim calculates phones' CM using the likelihood between phoneme model and anti-model, and then word's CM is obtained by averaging phone-level CMs[1]. In conventional method, CMs of some specific keywords are tory low and they are usually rejected. The reason is that statistics of phone-level CMs are not consistent. In other words, phone-level CMs have different probability density functions (pdf) for each phone, especially sri-phone. To overcome this problem, in this paper, we propose normalized confidence measure. Our approach is to transform CM pdf of each tri-phone to the same pdf under the assumption that CM pdfs are Gaussian. For evaluating our method we use common keyword spotting system. In that system context-dependent HMM models are used for modeling keyword utterance and contort-independent HMM models are applied to non-keyword utterance. The experiment results show that the proposed NCM reduced FAR (false alarm rate) from 0.44 to 0.33 FA/KW/HR (false alarm/keyword/hour) when MDR is about 8%. It achieves 25% improvement of FAR.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2024.05a
• /
• pp.774-776
• /
• 2024

The effective implementation of advanced speech recognition (ASR) systems necessitates the deployment of sophisticated keyword spotting models that are both responsive and resource-efficient. The initial local detection of user interactions is crucial as it allows for the selective transmission of audio data to cloud services, thereby reducing operational costs and mitigating privacy risks associated with continuous data streaming. In this paper, we address these needs and propose utilizing the Whisper-Tiny model with fine-tuning process to specifically recognize keywords from google speech dataset which includes 65000 audio clips of keyword commands. By adapting the model's encoder and appending a lightweight classification head, we ensure that it operates within the limited resource constraints of local devices. The proposed model achieves the notable test accuracy of 92.94%. This architecture demonstrates the efficiency as on-device model with stringent resources leading to enhanced accessibility in everyday speech recognition applications.

• Journal of Korea Multimedia Society
• /
• v.11 no.11
• /
• pp.1471-1479
• /
• 2008

In this paper, we propose a technique to improve Out-Of-Vocabulary(OOV) rejection algorithms in variable vocabulary recognition system which is much used in ASR(Automatic Speech Recognition). The rejection system can be classified into two categories by their implementation method, keyword spotting method and utterance verification method. The utterance verification method uses the likelihood ratio of each phoneme Viterbi score relative to anti-phoneme score for deciding OOV. In this paper, we add speaker verification system before utterance verification and calculate an speaker verification probability. The obtained speaker verification probability is applied for determining the proposed variable-confidence threshold. Using the proposed method, we achieve the significant performance improvement; CA(Correctly Accepted for keyword) 94.23%, CR(Correctly Rejected for out-of-vocabulary) 95.11% in office environment, and CA 91.14%, CR 92.74% in noisy environment.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.7
• /
• pp.632-640
• /
• 2002

In this paper, we proposed ACM (Anti-filler confidence measure) to compensate shortcoming of conventional RLJ-CM (RLJ-CM) and NCM (normalized CM), and integrated proposed ACM and conventional NCM using HCM (hybrid CM). Proposed ACM analyzes that FA (false acceptance) happens by the construction method of anti-phone model, and presumed phoneme sequence in actuality using phoneme recognizer to compensate this. We defined this as anti-phone model and used in confidence measure calculation. Analyzing feature of two confidences measure, conventional NCM shows good performance to FR (false rejection) and proposed ACM shows good performance in FA. This shows that feature of each other are complementary. Use these feature, we integrated two confidence measures using weighting vector α And defined this as HCM. In MDR (missed detection rate) 10% neighborhood, HCM is 0.219 FA/KW/HR (false alarm/keyword/hour). This is that Performance improves 22% than used conventional NCM individually.

• Annual Conference on Human and Language Technology
• /
• 1996.10a
• /
• pp.153-157
• /
• 1996

본 논문은 연속 분포 HMM을 사용한 핵심어 추출기법(Keyword Spotting)과 화자 인식에 기반한 음성 다이얼링 및 부서 안내에 관한 것이다. 개발된 시스템은 상대방의 이름, 직책, 존칭 등에 감탄사나 명령어 등이 혼합된 형태의 자연스런 음성 문장으로부터 다이얼링과 안내에 필요한 핵심어를 자동 추출하고 있다. 핵심 단어의 사용에는 자연성을 고려하여 문법적 제약을 최소한으로 두었으며, 각 단어 모델에 대해서는 음소의 갯수 더하기 $3{\sim}4$개의 상태 수와 3개 정도의 mixture component로써 좌우향 모델을, 묵음모델에 대해서는 2개 상태의 ergodic형 모델을 구성하였다. 인식에 있어서는 프레임 동기 One-Pass 비터비 알고리즘과 beam pruning을 채택하였으며, 인식에 사용된 어휘는 36개의 성명, 8개의 직위 및 존칭, 5개 정도의 호출어, 부탁을 나타내는 동사 및 그 활용이 10개 정도이다. 약 $3{\sim}6$개 정도의 단어로 구성된 문장을 실시간($1{\sim}3$초이내)에 인식하고, 약 98% 정도의 핵심어 인식 성능을 나타내고 있다.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.3
• /
• pp.3-9
• /
• 1999

This paper presents the study on the rejection capability based on anti-phone modeling for vocabulary independent speech recognition system. The rejection system detects and rejects out-of-vocabulary words which were not included in candidate words which are defined while the speech recognizer is made. The rejection system can be classified into two categories by their implementation methods, keyword spotting method and utterance verification method. The keyword spotting method uses an extra filler model as a candidate word as well as keyword models. The utterance verification method uses the anti-models for each phoneme for the calculation of confidence score after it has constructed the anti-models for all phonemes. We implemented an utterance verification algorithm which can be used for vocabulary independent speech recognizer. We also compared three kinds of means for the calculation of confidence score, and found out that the geometric mean had shown the best result. For the normalization of confidence score, usually Sigmoid function is used. On using it, we compared the effect of the weight constant for Sigmoid function and determined the optimal value. And we compared the effects of the size of cohort set, the results showed that the larger set gave the better results. And finally we found out optimal confidence score threshold value. In case of using the threshold value, the overall recognition rate including rejection errors was about 76%. This results are going to be adapted for stock information system based on speech recognizer which is currently provided as an experimental service by Korea Telecom.

• Proceedings of the KIEE Conference
• /
• 2008.04a
• /
• pp.170-171
• /
• 2008

연속적인 음성 인식 결과는 띄어쓰기를 하지 않은 연속 음절 문장들로 이루어져 있다. 본 논문은 음성 인식 후처리 단계에서 연속 음절 문장을 조사/어미 사전을 이용한 어절 생성 과정과 형태소 분석기를 이용하여 어절을 생성한 후 키워드를 추출한다. 실험 결과, 어절 생성기만 적용한 방식보다 제안된 알고리즘의 인식률이 향상되는 것을 확인하였다.

• MALSORI
• /
• no.46
• /
• pp.103-115
• /
• 2003

In this paper, we proposed Background Model Set algorithm used in the speaker verification to improve calculating confidence measure(CM) in speech recognition. CM is to display relative likelihood between recognized models and antiphone models. In previous method calculating of CM, we calculated probability and standard deviation using all phonemes in composition of antiphone models. At this process, antiphone CM brought bad recognition result. Also, recognition time increases. In order to solve this problem, we studied about method to reconstitute average and standard deviation using BMS algorithm in CM calculation.

• Proceedings of the Acoustical Society of Korea Conference
• /
• autumn
• /
• pp.81-84
• /
• 2001

본 논문은 중규모 단어급의 핵심어 검출 시스템에서 인식률 향상을 위해 미등록어 거절(Out-of-Vocabulary rejection) 기능을 제어하기 위한 연구이다. 이것은 핵심어 검출기에서 인식된 결과를 확인하는 과정으로 검증시스템이 구현되기 위해서는 매 음소마다 검증 기능이 필요하고, 이를 위해서 반음소(anti-phoneme model) 모델을 사용하였다. 검증의 역할은 인식기에서 인식된 단어가 등록어인지 미등록어인지 판별하는 것이다. 단어인식기는 비터비 탐색을 하므로, 기본적으로 단어단위로 인식을 하지만 그 인식된 단어는 내부적으로 음소단위로 인식된다. 따라서, 최소 검증 오류를 갖는 반음소 모델을 사용하고, 이를 이용하여 인식된 음소 단위들을 각각의 반음소 모델과 비교하여 통계적인 방법에 의해 신뢰도를 구한다 이 음소단위의 신뢰도를 단어 단위의 신뢰도로 환산하기 위해서 음소단위를 평균 내는 방식 을 취한다. 이렇게 함으로서, 등록어와 미등록어 사이의 분별력을 크게 하여 향상된 인식 성능을 얻었다.