가변 어휘 음성 인식기 구현 및 탐색 시간 단축 알고리듬 비교

서봉수 전남대학교 대학원 2001 국내석사

음성은 인간이 사용하는 가장 간단하고 효율적인 정보교환 수단이다. 이러한 음성을 컴퓨터에 응용하는 음성 인식에 대한 연구는 끊임없이 계속 되어져 왔고 현재는 이를 응용한 제품들이 나와있다. 그러나 그러한 제품들은 인식 어휘의 제한과 사용자의 사전 학습을 필요로 하는 불편함이 있다. 본 논문에서는 대용량 어휘 기반으로 사용 어휘의 제한 없이 사용자가 인식 하고자 하는 어휘의 사전 학습 없이 단어 입력만으로 인식 시스템을 사용자의 목적에 맞게 사용 할 수 있도록 하였다. 이를 위하여 본 논문에서는 결정 트리기반 상태 공유 알고리듬을 이용하여 가변 어휘를 구현하였고 가우시안 셀렉션 알고리듬 그리고 ENNS 알고리듬과 스칼라 양자화를 이용하여 인식 속도 단축을 하였다. 가우시안 셀렉션 알고리듬은 threshold 0.6에서 인식 속도는 약 15% 정도 향상이 되었으나 인식률에서는 약 10%의 감소를 보였다. 스칼라 양자화를 이용한 빠른 탐색 알고리듬에서는 인식 속도가 약 10%에서 12% 감소하였으나 인식률은 1%의 감소를 보였다. 인식 속도 향상에서는 가우시안 셀렉션 알고리듬이 보다 나은 성능을 보였으나 인식률 저하가 더 많았다. Speech is most simplest and efficient information exchange method. The study of speech recognition using the computer have been continued and many applications of speech recognition was producted. But this applications have many restrictions - word restriction, pre-training , slow recognition etc. So we realized of variable word speech recognition using decision tree based state tying algorithm and used gaussian selection algorithm, ENNS algorithm and scalar quantization algorithm for reduce search time in large vocabulary recognition system. Using the decision tree based state tying algorithm, the user use the recognition system without pre-training of recognized word. Gaussian selection algorithm is reduced search time about 15 % but recognition rate degradation is about 10 % at threshold 0.4. Fast search algorithm using scalar quantization is reduced search time about 10 %∼12 % and degraded only 1 % recognition rate. The gaussian selection algorithm is more efficient than scalar quantization in the reduce search time, but the recognition rate is not better.

PDA 상에서 음성제어를 위한 가변어휘 인식기 설계

郭相勳 東新大學校 2002 국내석사

음성인석 기술의 응용분야는 광범위하다. 현재 음성인식 기술은 정보통신 기기의 이동성을 극대화하는 모바일 컴퓨팅으로 그 범위가 확산되고 있다. 특히 인터넷 시장의 폭발적인 성장으로 사용자들의 인터넷 환경에 대한 인식이 모바일 환경으로 발전하게 되었다. 이러한 상황에 사용자는 더욱 빨라진 데이터 전송속도와 작고 가벼운 휴대자치로 데이터에 더욱 편리하게 접근하고자하였고 이를 만족시키는 것이 바로 PDA이다. 따라서, 본 논문은 이러한 PDA의 특징을 극대화하기 위해 트라이폰 기반의 가변어휘 음성인식기를 설계하여 음성제어를 구현하였다. 가면어휘 인식기는 결정 트리 기반 상태 공유 알고리즘을 기반으로 설계되었으며, 인식실험을 위해 Stand-alone 모델과 PDA에 적용시켜 비교실험을 하였다. 그 결과 Stand-alone 모델이 PDA상에서의 실험보다 인식률과 인식속도 면에서 모두 우수한 결과를 나타내었다. 하지만 PDA 상에서의 가변어휘 음성인식기를 이용하여 음성을 제어함으로써 PDA의 궁극적 특징인 휴대성과 이동성 및 보편성을 극대화하였다. The technology of speech recognition has wide field of application. The range of sucg technology is spreading into mobile computing doing the greatest of a movement about communication equipments at the present time. Particularly, recognition of internet environment developed onto mobile environment. Because of these environments urers want the faster speed of data transmission and the lighter portable equipment, which is PDA, for data access. Therefore, I designed text independent vocabulary recognizer based on tri-phone for implementation of speech control in this theses. The text independent vocabulary recognizer is based on decision trees state share algorithm. I made a recognition experiment by being applied it to stand-alone model and PDA for recognition experiments. Stand-alone model is superior to PDA in the respect of recognition rate and velocity. But I make the greatest of portable, movement and universality, which is the ultimate characteristic of PDA, by controlling voice using text independent vocabulary ecognizer.

TMS320C6711을 利用한 可變語彙 音聲認識器 구현

최지혁 成均館大學校 大學院 2003 국내석사

This word variable speech recognition system constituted the male and female vocabulary independent speech recognition system used CV, VCCV, and CV recognition units. This proposed sysem is consisted of CV 308, VCCV 3308, and CV 56 reference models, that is extracted from on-syllable database, name database, PBW (Phonetically Balanced Word), and general word database of male and female speaker. This system developed to embedded system(Home Automation, Telematics, Robot, Electric Home Appliances) for device independent speech recognition system. Code of this system is made according to ANSI C rules. To reduce all models size, we appled companding method to CV, VCCV and VC model probability. so we could reduce the previous 14.5Mbye(biniary file type)model memory to 3.5Mbyte(biniary file type). Therefor, Total size of recognliaer composed of under the 4Mbyte. This system efficiently canbe applied to a variable embedded system.

한국어 가변어휘 인식기의 성능향상에 관한 연구

정용원 부산대학교 일반대학원 2000 국내석사

In this paper, acoustic modeling and OOV rejection method were studied for Korean vocabulary-independent speech recognizer. To accurately model the phoneme, triphone was used and state-tying method was introduced for robust modeling with limited speech corpus. The problem of unseen model which appears in recognition phase but not in training phase was solved with Tree-Based Clustering that is one of top-down methodologies. In TBC, several phonetic question sets were organized and the best recognition result was achieved with question set that includes versatile phonetic question and excludes monophone-based question. Therefore, phonetic question set for TBC must include various phonetic phenomena and doesn't have to include monophone-based question. By measuring the confidence of recognized result, OOV rejection experiment was conducted. Two different methods were compared. One was based on utterance-level LLR and the other was based on frame-level LLR. For utterance-level OOV rejection experiment, best and 2^(nd) best result were used to get LLR. By normalizing the result with the length of utterance, better result was obtained. In comparison to the utterance-level OOV rejection, frame-level OOV rejection showed the better performance. In frame-level OOV rejection, filler model made from CI models was used for alternate hypothesis and the number of clusters that constitute filler model was varied. With filler model composed of two clusters, EER of 0.5% was achieved. This amounts to rejecting one IV word and accepting one OOV word out of 200 words. For future study, a novel method should be investigated for improved acoustic modeling. And for the OOV rejection, anti-modeling based on discriminative training method also should be tried. Additionally, for real-field application noise processing and keyword spotting also have to be implemented.

改善된 VCCV單位 HMM을 利用한 話者獨立 可變語彙 音聲認識 system

김우성 成均館大學校 大學院 2002 국내석사

In this paper, we implement an improved vocabulary-independent speech recognition system that has various states VCCV; it uses CV, VCCV, and VC recognition units. According to the phonetical characteristics, we statistically choose the best VCCV model instead of the fixed VCCV state model. If there is not the VCCV recognition unit, we can make a VCCV model which composes of the VC and the CV semi-syllable model. Applying the improved VCCV recognition unit to two words sets, the word recognition accuracy of the first words set is 96.4%, that of the second words set is 94.4% for male and 93.2%, 86.4% respectively for female. Comparing the improved VCCV recognition model(3∼7state) with the fixed VCCV(7state), we got the proposed recognition system which reduced 1.2% of the recognition errors in the first words set and 0.4% of the recognition errors in the second words set for male and 1.7%, 1.4% respectively for female. For male, the improved vocabulary-independent speech recognition system spent average 0.816184 seconds on recognizing a word of the first words set, which reduced 2.8% of the recognition time. And this system also spent average 0.706132 seconds on recognizing a word of the second words set, which reduced 2.3% of the recognition time. For female, this system spent average 0.88722 seconds on recognizing a word of the first words set, which reduced 1.62% of the recognition time. That also spent average 0.72156 seconds on recognizing a word of the second words set, which reduced 3.17% of the recognition time. For reducing all models size, we apply companding method to CV, VCCV, VC and silence model probability. so we reduce the previous 135Mbyte model memory(CV, VCCV, VC and silence HMM) to 12.7Mbyte. Using B model with reduced model memory, we got 91% word recognition accuracy of the first words set for male.

가변어휘 인식기의 계산량 감축을 위한 후보단어 사전검색

김영국 부산대학교 2001 국내석사

가변어휘 핵심어 검출 시스템의 구현 및 성능개선

신영욱 부산대학교 2001 국내석사