• The Journal of the Acoustical Society of Korea
• /
• v.35 no.2
• /
• pp.134-142
• /
• 2016

In this paper, a method synthesizing speech signal using the 40 kHz ultrasonic signals reflected from the articulatory muscles was introduced and performance was evaluated. When the ultrasound signals are radiated to articulating face, the Doppler effects caused by movements of lips, jaw, and chin observed. The signals that have different frequencies from that of the transmitted signals are found in the received signals. These ADS (Acoustic-Doppler Signals) were used for estimating of the speech parameters in this study. Prior to synthesizing speech signal, a quantitative correlation analysis between ADS and speech signals was carried out on each frequency bin. According to the results, the feasibility of the ADS-based speech synthesis was validated. ADS-to-speech transformation was achieved by the joint Gaussian mixture model-based conversion rules. The experimental results from the 5 subjects showed that filter bank energy and LPC (Linear Predictive Coefficient) cepstrum coefficients are the optimal features for ADS, and speech, respectively. In the subjective evaluation where synthesized speech signals were obtained using the excitation sources extracted from original speech signals, it was confirmed that the ADS-to-speech conversion method yielded 72.2 % average recognition rates.

• Journal of the Korea Society of Computer and Information
• /
• v.24 no.12
• /
• pp.17-24
• /
• 2019

The biggest problem in the AI field, which has become a hot topic in recent years, is how to deal with the lack of training data. Since manual data construction takes a lot of time and efforts, it is non-trivial for an individual to easily build the necessary data. On the other hand, automatic data construction needs to handle data quality issue. In this paper, we introduce a method to automatically extract the data required to develop Korean speech command recognizer from the web and to automatically select the data that can be used for training data. In particular, we propose a modified ResNet model that shows modest performance for the automatically constructed Korean speech command data. We conducted an experiment to show the applicability of the command set of the health and daily life domain. In a series of experiments using only automatically constructed data, the accuracy of the health domain was 89.5% in ResNet15 and 82% in ResNet8 in the daily lives domain, respectively.

• Phonetics and Speech Sciences
• /
• v.12 no.3
• /
• pp.55-63
• /
• 2020

Voice conversion can be applied to various voice processing applications. It can also play an important role in data augmentation for speech recognition. The conventional method uses the architecture of voice conversion with speech synthesis, with Mel filter bank as the main parameter. Mel filter bank is well-suited for quick computation of neural networks but cannot be converted into a high-quality waveform without the aid of a vocoder. Further, it is not effective in terms of obtaining data for speech recognition. In this paper, we focus on performing voice-to-voice conversion using only the raw spectrum. We propose a deep learning model based on the transformer network, which quickly learns the voice conversion properties using an attention mechanism between source and target spectral components. The experiments were performed on TIDIGITS data, a series of numbers spoken by an English speaker. The conversion voices were evaluated for naturalness and similarity using mean opinion score (MOS) obtained from 30 participants. Our final results yielded 3.52±0.22 for naturalness and 3.89±0.19 for similarity.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.39 no.2
• /
• pp.116-127
• /
• 2002

A new approach to improve the speaker adaptation algorithm by means of the variable number of observation density functions for CDHMM speech recognizer has been proposed. The proposed method uses the observation density function with more than one mixture in each state to represent speech characteristics in detail. The number of mixtures in each state is determined by the number of frames and the determinant of the variance, respectively. The each MAP Parameter is extracted in every mixture determined by these two methods. In addition, the state segmentation method requiring speaker adaptation can segment the adapting speech more Precisely by using speaker-independent model trained from sufficient database as a priori knowledge. And the state duration distribution is used lot adapting the speech duration information owing to speaker's utterance habit and speed. The recognition rate of the proposed methods are significantly higher than that of the conventional method using one mixture in each state.

• Journal of the Korea Computer Industry Society
• /
• v.5 no.4
• /
• pp.491-502
• /
• 2004

In this paper, we propose a HMM model using FSVQ(First Section VQ), fuzzy theory and doubly spectral feature, as study on the isolated word recognition system of speaker-independent. In the proposed paper, LPC cepstrum coefficients and regression coefficients of LPC cepstrum as doubly spectral feature be used. And, training data are divided several section and first section is generated codebook of VQ, and then is obtained multi-observation sequences by order of large propabilistic values based on fuzzy nile from the codebook of the first section. Thereafter, this observation sequences of first section is trained and is recognized a word to be obtained highest probaility by same concept. Besides the speech recognition experiments of proposed method, we experiment the other methods under the equivalent environment of data and conditions. In the whole experiment, it is proved that the proposed method is superior to the others in recognition rate.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.2
• /
• pp.11-17
• /
• 1999

This paper proposes an observation probability smoothing technique for the robustness of a discrete hidden Markov(DHMM) model based speech recognizer. Also, an appropriate noise robust processing in car environment is suggested from experimental results. The noisy speech is often mislabeled during the vector quantization process. To reduce the effects of such mislabelings, the proposed technique increases the observation probability of similar codewords. For the noise robust processing in car environment, the liftering on the distance measure of feature vectors, the high pass filtering, and the spectral subtraction methods are examined. Recognition experiments on the 14-isolated words consists of the Korean digits and command words were performed. The database was recorded in a stopping car and a running car environments. The recognition rates of the baseline recognizer were 97.4% in a stopping situation and 59.1% in a running situation. Using the proposed observation probability smoothing technique, the liftering, the high pass filtering, and the spectral subtraction the recognition rates were enhanced to 98.3% in a stopping situation and to 88.6% in a running situation.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.18 no.7
• /
• pp.1740-1748
• /
• 2014

This paper focuses on listening test and acoustic analysis of patients' speech with velopharyngeal insufficiency (VPI) and normal speakers' simulation speech. In this research, a set consisting of 50-words, vowels and single syllables is determined for speech database construction. A web-based listening evaluation system is developed for a convenient/automated evaluation procedure. The analysis results show the trend of incorrect recognition for VPI speech and the one for simulation speech are similar. Such similarity is also confirmed by comparing the formant locations of vowel and spectrum of consonant sounds. These results show that the simulation method for VPI speech is effective at generating the speech signals similar to actual VPI patient's speech. It is expected that the simulation speech data can be effectively employed for our future work such as acoustic model adaptation.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.461-467
• /
• 2020

In noisy and multi-speaker environments, the performance of speech recognition is unavoidably lower than in a clean environment. To improve speech recognition, in this paper, the signal of the speaker of interest is extracted from the mixed speech signals with multiple speakers. The VoiceFilter model is used to effectively separate overlapped speech signals. In this work, clustering by Probabilistic Linear Discriminant Analysis (PLDA) similarity score was employed to detect the speech signal of the interested speaker, which is used as the reference speaker to VoiceFilter-based separation. Therefore, by utilizing the speaker feature extracted from the detected speech by the proposed clustering method, this paper propose a speaker diarization system using only the mixed speech without an explicit reference speaker signal. We use phone-dataset consisting of two speakers to evaluate the performance of the speaker diarization system. Source to Distortion Ratio (SDR) of the operator (Rx) speech and customer speech (Tx) are 5.22 dB and -5.22 dB respectively before separation, and the results of the proposed separation system show 11.26 dB and 8.53 dB respectively.

• 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.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.32B no.5
• /
• pp.765-776
• /
• 1995

This paper describes a speaker adaptation method to improve the recognition performance of MLP(multiLayer Perceptron) based HMM(Hidden Markov Model) speech recognizer. In this method, we use lst-order linear transformation network to fit data of a new speaker to the MLP. Transformation parameters are adjusted by back-propagating classification error to the transformation network while leaving the MLP classifier fixed. The recognition system is based on semicontinuous HMM's which use the MLP as a fuzzy vector quantizer. The experimental results show that rapid speaker adaptation resulting in high recognition performance can be accomplished by this method. Namely, for supervised adaptation, the error rate is signifecantly reduced from 9.2% for the baseline system to 5.6% after speaker adaptation. And for unsupervised adaptation, the error rate is reduced to 5.1%, without any information from new speakers.