• The Journal of the Acoustical Society of Korea
• /
• v.31 no.4
• /
• pp.266-272
• /
• 2012

Speaker segmentation is a process of automatically detecting the speaker boundary points in the audio data. Speaker segmentation methods are divided into two categories depending on whether they use a prior knowledge or not: One is the model-based segmentation and the other is the metric-based segmentation. In this paper, we introduce the eigenvoice-based speaker weight distance method and compare it with the representative metric-based methods. Also, we employ and compare the Euclidean and cosine similarity functions to calculate the distance between speaker weight vectors. And we verify that the speaker weight distance method is computationally very efficient compared with the method directly using the distance between the speaker adapted models constructed by the eigenvoice technique.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.497-504
• /
• 2020

Constructing speaker embeddings in speaker verification is an important issue. In general, a self-attention mechanism has been applied for speaker embedding encoding. Previous studies focused on training the self-attention in a high-level layer, such as the last pooling layer. In this case, the effect of low-level layers is not well represented in the speaker embedding encoding. In this study, we propose Masked Cross Self-Attentive Encoding (MCSAE) using ResNet. It focuses on training the features of both high-level and low-level layers. Based on multi-layer aggregation, the output features of each residual layer are used for the MCSAE. In the MCSAE, the interdependence of each input features is trained by cross self-attention module. A random masking regularization module is also applied to prevent overfitting problem. The MCSAE enhances the weight of frames representing the speaker information. Then, the output features are concatenated and encoded in the speaker embedding. Therefore, a more informative speaker embedding is encoded by using the MCSAE. The experimental results showed an equal error rate of 2.63 % using the VoxCeleb1 evaluation dataset. It improved performance compared with the previous self-attentive encoding and state-of-the-art methods.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.466-472
• /
• 2021

Speaker Change Detection (SCD) refers to finding the moment when the main speaker changes from one person to the next in a speech conversation. In speaker change detection, difficulties arise due to overlapping speakers, inaccuracy in the information labeling, and data imbalance. To solve these problems, TIMIT corpus widely used in speech recognition have been concatenated artificially to obtain a sufficient amount of training data, and the detection of changing speaker has performed after identifying overlapping speakers. In this paper, we propose an speaker change detection system that considers the speaker overlapping. We evaluated and verified the performance using various approaches. As a result, a detection system similar to the X-Vector structure was proposed to remove the speaker overlapping region, while the Bi-LSTM method was selected to model the speaker change system. The experimental results show a relative performance improvement of 4.6 % and 13.8 % respectively, compared to the baseline system. Additionally, we determined that a robust speaker change detection system can be built by conducting related studies based on the experimental results, taking into consideration text and speaker information.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.2
• /
• pp.137-142
• /
• 2020

Many studies based on I-vector have been conducted in a variety of environments, from text-dependent short-utterance to text-independent long-utterance. In this paper, we propose a speaker verification system employing a combination of I-vector with Probabilistic Linear Discriminant Analysis (PLDA) and speaker embedding of Long Short Term Memory (LSTM) with attention mechanism in far-field and noisy environments. The LSTM model's Equal Error Rate (EER) is 15.52 % and the Attention-LSTM model is 8.46 %, improving by 7.06 %. We show that the proposed method solves the problem of the existing extraction process which defines embedding as a heuristic. The EER of the I-vector/PLDA without combining is 6.18 % that shows the best performance. And combined with attention-LSTM based embedding is 2.57 % that is 3.61 % less than the baseline system, and which improves performance by 58.41 %.

• Annual Conference on Human and Language Technology
• /
• 2012.10a
• /
• pp.77-83
• /
• 2012

본 연구에서는 로봇의 자동 동화구연을 목표로 발화문장 상의 감정 파악 및 등장인물 별 다앙한 TTS 보이스 선택에 활용 가능한 발화문장의 화자 파악문제를 다룬다. 본 연구에서는 기존 규칙기반 방법론에서 많이 활용되어온 자질인 후보의 위치, 화자 후보의 주격/목적격 여부, 발화동사 존재 여부를 비롯하여 동화에 자주 나타나는 등장인물의 의미적 분류 및 등장인물의 등장/퇴장과 관련된 동사들을 추가 자질로 활용한다. 사람 및 동식물, 무생물이 모두 화자가 될 수 있는 동화 코퍼스에서 제안한 자질들을 활용하여 의사결정트리로 학습 및 검증한 결과 규칙기반의 베이스라인 방법에 비해 최대 49%의 정확도가 향상되었고, 제안한 방법론이 데이터의 변화에도 강인한 것을 확인할 수 있었다.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.19 no.6
• /
• pp.755-759
• /
• 2009

The ability to recognize human emotion is one of the hallmarks of human-robot interaction. Especially, speaker-independent emotion recognition is a challenging issue for commercial use of speech emotion recognition systems. In general, speaker-independent systems show a lower accuracy rate compared with speaker-dependent systems, as emotional feature values depend on the speaker and his/her gender. Hence, this paper describes the realization of speaker-independent emotion recognition by rejection using confidence measure to make the emotion recognition system be homogeneous and accurate. From comparison of the proposed methods with conventional method, the improvement and effectiveness of proposed methods were clearly confirmed.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.12 no.5
• /
• pp.2402-2409
• /
• 2011

The voice has distinguishable acoustic properties of speaker as well as transmitting information. The speaker recognition is the method to figures out who speaks the words through acoustic differences between speakers. The speaker recognition is roughly divided two kinds of categories: speaker verification and identification. The speaker verification is the method which verifies speaker himself based on only one's voice. Otherwise, the speaker identification is the method to find speaker by searching most similar model in the database previously consisted of multiple subordinate sentences. This paper composes feature vector from extracting MFCC coefficients and uses the dynamic time warping algorithm to compare the similarity between features. In order to describe common characteristic based on phonological features of spoken words, two subordinate sentences for each speaker are used as the training data. Thus, it is possible to identify the speaker who didn't say the same word which is previously stored in the database.

• Journal of the Korean Institute of Telematics and Electronics T
• /
• v.36T no.4
• /
• pp.17-23
• /
• 1999

This paper described a security system using text-independent speaker recognition algorithm. Security system is based on PIC16F84 and sound card. Speaker recognition algorithm applied a k-means based model and weighted cepstrum for speech features. As the experimental results, recognition rate of the training data is 100%, non-training data is 99%. Also false rejection rate is 1%, false acceptance rate is 0% and verification mean error rate is 0.5% for registered 5 persons.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.6
• /
• pp.26-33
• /
• 2001

Though speaker recognition based on conventional AR HMM shows good performance, its lack of modeling the environmental noise makes its performance degraded in case of practical noisy environment. In this paper, a robust speaker recognition system based on AR HMM is proposed, where noise is considered in the observation signal model for practical noisy environment and duration-term is considered to increase performance. Experimental results, using the digits database from 100 speakers (77 males and 23 females) under white noise and car noise, show improved performance.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.4
• /
• pp.299-307
• /
• 2003

There have been many studies on speaker normalization which aims to minimize the effects of speaker's vocal tract length on the recognition performance of the speaker independent speech recognition system. In this paper, we propose a simple vector quantizer based linear warping speaker normalization method based on the observation that the vector quantizer can be successfully used for speaker verification. For this purpose, we firstly generate an optimal codebook which will be used as the basis of the speaker normalization, and then the warping factor of the unknown speaker will be extracted by comparing the feature vectors and the codebook. Finally, the extracted warping factor is used to linearly warp the Mel scale filter bank adopted in the course of MFCC calculation. To test the performance of the proposed method, a series of recognition experiments are conducted on discrete HMM with thirteen mono-syllabic Korean number utterances. The results showed that about 29% of word error rate can be reduced, and that the proposed warping factor extraction method is useful due to its simplicity compared to other line search warping methods.