• Phonetics and Speech Sciences
• /
• v.6 no.1
• /
• pp.77-83
• /
• 2014

This paper proposes a speaker and environment clustering method in order to overcome the degradation of the speech recognition performance caused by various noise and speaker characteristics. In this paper, instead of using the distance between Gaussian mixture model (GMM) weight vectors as in the Google's approach, the distance between the adapted mean vectors based on the modified maximum a posteriori (MAP) adaptation is used as a distance measure for vector quantization (VQ) clustering. According to our experiments on the simulation data generated by adding noise to clean speech, the proposed clustering method yields error rate reduction of 10.6% compared with baseline speaker-independent (SI) model, which is slightly better performance than the Google's approach.

• Journal of Broadcast Engineering
• /
• v.23 no.4
• /
• pp.511-518
• /
• 2018

This paper deals with improving speech intelligibility by applying binary mask to time-frequency units of speech in noise. The binary mask is set to "0" or "1" according to whether speech is dominant or noise is dominant by comparing signal-to-noise ratio with pre-defined threshold. Bayesian classifier trained with Gaussian mixture model is used to estimate the binary mask of each time-frequency signal. The binary mask based noise suppressor improves speech intelligibility only in noise condition which is included in the training data. In this paper, speaker adaptation techniques for speech recognition are applied to adapt the Gaussian mixture model to a new noise environment. Experiments with noise-corrupted speech are conducted to demonstrate the improvement of speech intelligibility by employing adaption techniques in a new noise environment.

• Phonetics and Speech Sciences
• /
• v.10 no.3
• /
• pp.41-48
• /
• 2018

This study examined the acoustic characteristics of Korean alveolar fricatives of normal children. Developing children aged 3 and 7, typically produced 2 types of nonsense syllables containing alveolar fricative /sV/ and /VsV/ sequences where V was any one of three corner vowels (/i, a, and u/). Stimuli containing the speech materials used in a production experiment were presented randomly to 12 speech language pathologists (SLPs) for a perception test. The SLPs responded by selecting one of seven alternative sounds. Acoustic measures such as duration of frication noise, normalized intensity, skewness, and center of gravity were examined. There was significant difference in acoustic measures when comparing vowels. Comparison of syllable structures indicated statistically significant differences in duration of frication noise and normalized intensity. Acoustic parameters could account for the perceptual data. Relating the acoustic and perception data by means of logistic regression suggests that duration of frication noise and normalized intensity are the primary cues to perceiving Korean fricatives.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.2
• /
• pp.21-26
• /
• 2001

An expectation-maximization (EM) based Bayesian adaptation method for the mean of noise is proposed for noise-robust speech recognition. In the algorithm, the on-line testing utterances are used for the unsupervised Bayesian adaptation and the prior distribution of the noise mean is estimated using the off-line training data. For the noisy speech modeling, the parallel model combination (PMC) method is employed. The proposed method has shown to be effective compared with the conventional PMC method for the speech recognition experiments in a car-noise condition.

• Journal of Broadcast Engineering
• /
• v.18 no.2
• /
• pp.311-318
• /
• 2013

This paper deals with a noise reduction algorithm which uses the binary masking in the time-frequency domain. To improve speech intelligibility in noise, noise-masked speech is decomposed into time-frequency units and mask "0" is assigned to masker-dominant region removing time-frequency units where noise is dominant compared to speech. In the previous research, Gaussian mixture models were used to classify the speech-dominant region and noise-dominant region which correspond to mask "1" and mask "0", respectively. In each frequency band, data were collected and trained to build the Gaussian mixture models and detection procedure is performed to the test data where each time-frequency unit belongs to speech-dominant region or noise-dominant region. In this paper, we consider the correlation of masks in the frequency domain and propose a post-processing method which exploits the Viterbi algorithm.

• Journal of Convergence for Information Technology
• /
• v.11 no.6
• /
• pp.1-6
• /
• 2021

In the field of speech recognition, as the DNN is applied, the use of speech recognition is increasing, but the amount of calculation for parallel training needs to be larger than that of the conventional GMM, and if the amount of data is small, overfitting occurs. To solve this problem, we propose an efficient method for robust voice feature extraction and voice signal noise removal even when the amount of data is small. Speech feature extraction efficiently extracts speech energy by applying the difference in frame energy for speech and the zero-crossing ratio and level-crossing ratio that are affected by the speech signal. In addition, in order to remove noise, the noise of the speech signal is removed by removing the noise of the speech signal with an average predictive improved LMS filter with little loss of speech information while maintaining the intrinsic characteristics of speech in detection of the speech signal. The improved LMS filter uses a method of processing noise on the input speech signal by adjusting the active parameter threshold for the input signal. As a result of comparing the method proposed in this paper with the conventional frame energy method, it was confirmed that the error rate at the start point of speech is 7% and the error rate at the end point is improved by 11%.

• Proceedings of the Korean Information Science Society Conference
• /
• 2006.10d
• /
• pp.302-305
• /
• 2006

In this paper, we have proposed a novel method using quantization of wavelet coefficients for secret data communication. First, speech signal is partitioned into small time frames and the frames are transformed into frequency domain using a WT(Wavelet Transform). We quantize the wavelet coefficients and embedded secret data into the quantized wavelet coefficients. The destination regard quantization errors of received speech as seceret dat. As most speech watermark techniques have a trade off between noise robustness and speech quality, our method also have. However we solve the problem with a partial quantization and a noise level dependent threshold. In additional, we improve the speech quality with de-noising method using wavelet transform. Since the signal is processed in the wavelet domain, we can easily adapt the de-noising method based on wavelet transform. Simulation results in the various noisy environments show that the proposed method is reliable for secret communication.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.5 no.4
• /
• pp.357-362
• /
• 2010

This paper proposes the speech and noise recognition system by using a neural network in order to detect the speech and noise sections at each frame. The proposed neural network consists of a layered neural network training by back-propagation algorithm. First, a power spectrum obtained by fast Fourier transform and linear predictive coefficients are used as the input to the neural network for each frame, then the neural network is trained using these power spectrum and linear predictive coefficients. Therefore, the proposed neural network can train using clean speech and noise. The performance of the proposed recognition system was evaluated based on the recognition rate using various speeches and white, printer, road, and car noises. In this experiment, the recognition rates were 92% or more for such speech and noise when training data and evaluation data were the different.

• The Journal of the Acoustical Society of Korea
• /
• v.41 no.3
• /
• pp.342-350
• /
• 2022

In this paper, actual drone noise samples are collected for speech processing in disaster environments to build noise-corrupted speech database, and speech enhancement performance is evaluated by applying spectrum subtraction and mask-based speech enhancement techniques. To improve the performance of VoiceFilter (VF), an existing deep neural network-based speech enhancement model, we apply the Self-Attention operation and use the estimated noise information as input to the Attention model. Compared to existing VF model techniques, the experimental results show 3.77%, 1.66% and 0.32% improvements for Source to Distortion Ratio (SDR), Perceptual Evaluation of Speech Quality (PESQ), and Short-Time Objective Intelligence (STOI), respectively. When trained with a 75% mix of speech data with drone sounds collected from the Internet, the relative performance drop rates for SDR, PESQ, and STOI are 3.18%, 2.79% and 0.96%, respectively, compared to using only actual drone noise. This confirms that data similar to real data can be collected and effectively used for model training for speech enhancement in environments where real data is difficult to obtain.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.3
• /
• pp.56-62
• /
• 2001

In this paper, we adopt a missing data theory to speech recognition. It can be used in order to maintain high performance of speech recognizer when the missing data occurs. In general, hidden Markov model (HMM) is used as a stochastic classifier for speech recognition task. Acoustic events are represented by continuous probability density function in continuous density HMM(CDHMM). The missing data theory has an advantage that can be easily applicable to this CDHMM. A marginalization method is used for processing missing data because it has small complexity and is easy to apply to automatic speech recognition (ASR). Also, a spectral subtraction is used for detecting missing data. If the difference between the energy of speech and that of background noise is below given threshold value, we determine that missing has occurred. We propose a new method that examines the reliability of detected missing data using voicing probability. The voicing probability is used to find voiced frames. It is used to process the missing data in voiced region that has more redundant information than consonants. The experimental results showed that our method improves performance than baseline system that uses spectral subtraction method only. In 452 words isolated word recognition experiment, the proposed method using the voicing probability reduced the average word error rate by 12％ in a typical noise situation.