• ETRI Journal
• /
• v.45 no.4
• /
• pp.678-689
• /
• 2023

This paper describes a novel end-to-end deep generative model-based speaker recognition system using prosodic features. The usefulness of variational autoencoders (VAE) in learning the speaker-specific prosody representations for the speaker recognition task is examined herein for the first time. The speech signal is first automatically segmented into syllable-like units using vowel onset points (VOP) and energy valleys. Prosodic features, such as the dynamics of duration, energy, and fundamental frequency (F₀), are then extracted at the syllable level and used to train/adapt a speaker-dependent VAE from a universal VAE. The initial comparative studies on VAEs and traditional autoencoders (AE) suggest that the former can efficiently learn speaker representations. Investigations on the impact of gender information in speaker recognition also point out that gender-dependent impostor banks lead to higher accuracies. Finally, the evaluation on the NIST SRE 2010 dataset demonstrates the usefulness of the proposed approach for speaker recognition.

• Journal of the Korea Society of Computer and Information
• /
• v.19 no.11
• /
• pp.159-166
• /
• 2014

In this paper, we examine the role of emotional acoustic cues including both prosody and voice quality parameters for the modification of a word sense. For the extraction of prosody parameters and voice quality parameters, we used 60 pieces of speech data spoken by six speakers with five different emotional states. We analyzed eight different emotional acoustic cues, and used a discriminant analysis technique in order to find the dominant sequence of acoustic cues. As a result, we found that anger has a close relation with intensity level and 2nd formant bandwidth range; joy has a relative relation with the position of 2nd and 3rd formant values and intensity level; sadness has a strong relation only with prosody cues such as intensity level and pitch level; and fear has a relation with pitch level and 2nd formant value with its bandwidth range. These findings can be used as the guideline for find-tuning an emotional spoken language generation system, because these distinct sequences of acoustic cues reveal the subtle characteristics of each emotional state.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.3E
• /
• pp.85-90
• /
• 2004

This paper presents two kinds of new approaches, one of which is concerned with recognition of emotional speech such as anger, happiness, normal, sadness, or surprise. The other is concerned with emotion recognition in speech. For the proposed speech recognition system handling human speech with emotional states, total nine kinds of prosodic features were first extracted and then given to prosodic identifier. In evaluation, the recognition results on emotional speech showed that the rates using proposed method increased more greatly than the existing speech recognizer. For recognition of emotion, on the other hands, four kinds of prosodic parameters such as pitch, energy, and their derivatives were proposed, that were then trained by discrete duration continuous hidden Markov models(DDCHMM) for recognition. In this approach, the emotional models were adapted by specific speaker's speech, using maximum a posteriori(MAP) estimation. In evaluation, the recognition results on emotional states showed that the rates on the vocal emotions gradually increased with an increase of adaptation sample number.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.333-335
• /
• 2022

The development of speech recognition processing technology has been applied and used in various video and streaming services along with STT and TTS technologies. However, there are high barriers to clear written expression due to the use of dialects and overlapping of stop words, exclamations, and similar words for voice recognition of actual conversation content. In this study, for ambiguous dialects in speech recognition, we propose a speech recognition technology that applies dialect key word dictionary processing method by category and dialect prosody as speech recognition network model properties.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2019.11a
• /
• pp.137-139
• /
• 2019

본 논문은 다양한 음성 특징과 텍스트를 이용한 멀티 모드 순환신경망 네트워크를 사용하여 음성을 통한 범주형(categorical) 분류 방법과 Arousal-Valence(AV) 도메인에서의 분류방법을 통해 감정인식 결과를 제시한다. 본 연구에서는 음성 특징으로는 MFCC, Energy, Velocity, Acceleration, Prosody 및 Mel Spectrogram 등의 다양한 특징들의 조합을 이용하였고 이에 해당하는 텍스트 정보를 순환신경망 기반 네트워크를 통해 융합하여 범주형 분류 방법과 과 AV 도메인에서의 분류 방법을 이용해 감정을 이산적으로 분류하였다. 실험 결과, 음성 특징의 조합으로 MFCC Energy, Velocity, Acceleration 각 13 차원과 35 차원의 Prosody 의 조합을 사용하였을 때 범주형 분류 방법에서는 75%로 다른 특징 조합들 보다 높은 결과를 보였고 AV 도메인 에서도 같은 음성 특징의 조합이 Arousal 55.3%, Valence 53.1%로 각각 가장 높은 결과를 보였다.

• Proceedings of the KSPS conference
• /
• 2003.10a
• /
• pp.175-179
• /
• 2003

In order to produce a high quality (intelligibility and naturalness) synthesized speech, it is very important to get an accurate grapheme-to-phoneme conversion and prosody model. In this paper, we analyzed Chinese texts using a segmentation, POS tagging and unknown word recognition. We present a grapheme-to-phoneme conversion using a dictionary-based and rule-based method. We constructed a prosody model using a probabilistic method and a decision tree-based error correction method. According to the result from the above analysis, we can successfully select and concatenate exact synthesis unit of syllables from the Chinese Synthesis DB.

• Speech Sciences
• /
• v.2
• /
• pp.159-172
• /
• 1997

This paper proposes a prosodic representation system of the Kyungnam dialect of Korean, based on the ToBI system. In this system, diverse intonation patterns are transcribed on the four parallel tiers: a tone tier, a break index tier, an orthographic tier, and a miscellaneous tier. The tone tier employs pitch accents, phrase accents, and boundary tones marked with diacritics in order to represent various pitch events. The break index tier uses five break indices, numbered from 0 to 4, in order to represent degrees of connectiveness in speech by associating each inter-word position with a break index. In this, each break index represents a boundary of some kind of constituent. This system can contribute not only to a more detailed theory connecting prosody, syntax, and intonation, but also to current text-to-speech synthesis approaches, speech recognition, and other quantitative computational modellings.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.2E
• /
• pp.98-104
• /
• 2002

The present study describes a combination method to recognize the human affective states such as anger, happiness, sadness, or surprise. For this, we extracted emotional features from voice signals and facial expressions, and then trained them to recognize emotional states using hidden Markov model (HMM) and neural network (NN). For voices, we used prosodic parameters such as pitch signals, energy, and their derivatives, which were then trained by HMM for recognition. For facial expressions, on the other hands, we used feature parameters extracted from thermal and visible images, and these feature parameters were then trained by NN for recognition. The recognition rates for the combined parameters obtained from voice and facial expressions showed better performance than any of two isolated sets of parameters. The simulation results were also compared with human questionnaire results.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.4E
• /
• pp.112-119
• /
• 2004

This paper presents a method for the automatic recognition of pitch accents with no prior knowledge about the phonetic content of the signal (no knowledge of word or phoneme boundaries or of phoneme labels). The recognition algorithm used in this paper is a time-delay recurrent neural network (TDRNN). A TDRNN is a neural network classier with two different representations of dynamic context: delayed input nodes allow the representation of an explicit trajectory F0(t), while recurrent nodes provide long-term context information that can be used to normalize the input F0 trajectory. Performance of the TDRNN is compared to the performance of a MLP (multi-layer perceptron) and an HMM (Hidden Markov Model) on the same task. The TDRNN shows the correct recognition of $91.9{\%}\;of\;pitch\;events\;and\;91.0{\%}$ of pitch non-events, for an average accuracy of $91.5{\%}$ over both pitch events and non-events. The MLP with contextual input exhibits $85.8{\%},\;85.5{\%},\;and\;85.6{\%}$ recognition accuracy respectively, while the HMM shows the correct recognition of $36.8{\%}\;of\;pitch\;events\;and\;87.3{\%}$ of pitch non-events, for an average accuracy of $62.2{\%}$ over both pitch events and non-events. These results suggest that the TDRNN architecture is useful for the automatic recognition of pitch accents.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1998.08a
• /
• pp.396-400
• /
• 1998

문장 독립형 화자인식 시스템에 운율정보 사용을 제안한다. 스펙트럴 특징패턴만을 주로 사용하고 있는 기존의 화자인식 시스템은 채널왜곡이나 기타 잡음환경에서 성능이 크게 저하된다. 그러나 화자의 speaking style을 반영하는 운율정보는 주위환경에 강인한 특성을 갖는다. 적합한 코드북 크기와 피치 컨투어 특징 벡터의 길이를 실험 치로 구하여 자동차 소음과 백색 가우시안 소음이 섞인 음성에 대하여 화자인식 실험을 하였다. 실험 결과 소음 환경에서 운율 정보를 이용한 화자 dsltlr 시스템이 스펙트럴 모델보다 인식율이 높음을 보였다.