• 한국음향학회:학술대회논문집
• /
• 한국음향학회 1985년도 학술발표회 논문집
• /
• pp.87-94
• /
• 1985

In this paper, enhancement of speech corrupted by additive white or colored noise is stuided. The nuconstrained frequency-domain block least-mean-square (UFBLMS) adaptation algorithm and its frequency-weighted version are newly applied to speech enhancement. For enhancement of speech degraded by white noise, the performance of the UFBLMS algorithm is superior to the spectral subtraction method or Wiener filtering technique by more than 3 dB in segmented frequency-weighted signal-to-noise ratio(FWSNERSEG) when SNR of speech is in the range of 0 to 10 dB. As for enhancement of noisy speech corrupted by colored noise, the UFBLMS algorithm is superior to that of the spectral subtraction method by about 3 to 5 dB in FWSNRSEG. Also, it yields better performance by about 2 dB in FWSNR and FWSNRSEG than that of time-domain least-mean-square (TLMS) adaptive prediction filter(APF). In view of the computational complexity and performance improvement in speech quality and intelligibility, the frequency-weighted UFBLMS algorithm appears to yield the best performance among various algorithms in enhancing noisy speech corrupted by white or colored noise.

• 한국멀티미디어학회논문지
• /
• 제22권7호
• /
• pp.749-758
• /
• 2019

This paper proposes a data augmentation algorithm to improve the performance of DNN(Deep Neural Network) based speech enhancement. Many deep learning models are exploring algorithms to maximize the performance in limited amount of data. The most commonly used algorithm is the data augmentation which is the technique artificially increases the amount of data. For the effective data augmentation algorithm, we used a formant enhancement method that assign the different weights to the formant frequencies. The DNN model which is trained using the proposed data augmentation algorithm was evaluated in various noise environments. The speech enhancement performance of the DNN model with the proposed data augmentation algorithm was compared with the algorithms which are the DNN model with the conventional data augmentation and without the data augmentation. As a result, the proposed data augmentation algorithm showed the higher speech enhancement performance than the other algorithms.

• 대한의용생체공학회:의공학회지
• /
• 제28권6호
• /
• pp.732-743
• /
• 2007

Background noise is one of the major complaints of not only hearing impaired persons but also normal listeners. This paper describes the results of two experiments in which speech recognition performance was determined for listeners with normal hearing and sensorineural hearing loss in noise environment. First, we compared speech enhancement algorithms by evaluation speech recognition ability in various speech-to-noise ratios and types of noise. Next, speech enhancement algorithms by reducing background noise were presented and evaluated to improve speech intelligibility for sensorineural hearing impairment listeners. We tested three noise reduction methods using single-microphone, such as spectrum subtraction and companding, Wiener filter method, and maximum likelihood envelop estimation. Their responses in background noise were investigated and compared with those by the speech enhancement algorithm that presented in this paper. The methods improved speech recognition test score for the sensorineural hearing impaired listeners, but not for normal listeners. The results suggest the speech enhancement algorithm with the loudness compression can improve speech intelligibility for listeners with sensorineural hearing loss.

• 한국전자통신학회논문지
• /
• 제16권3호
• /
• pp.465-470
• /
• 2021

잡음이 존재하는 실제 환경에서 음성인식을 실시하는 경우에 음성인식의 성능 열화 및 음성의 품질이 저화되지 않는 강건한 음성인식 기술이 필요하다. 이러한 음성인식 기술을 개발함으로써 사람의 음성 스펙트럼과 유사한 잡음 환경에서도 안정되고 높은 음성인식률이 실현되는 어플리케이션이 요구된다. 따라서 본 논문에서는 최소 평균 제곱의 오차를 기반으로 한 단시간 스펙트럼 진폭 방법인 MMSA-STSA 추정 알고리즘에 기초한 잡음억압을 처리하는 음성강조 알고리즘을 제안한다. 이 알고리즘은 단일 채널 입력에 기초한 효과적인 비선형 음성강조 알고리즘이며, 높은 잡음억제 성능을 가지고 있으며 음성의 통계적인 모델에 기초하여 음성의 왜곡량을 줄이는 기법이다. 본 실험에서는 MMSA-STSA 추정 알고리즘의 유효성을 확인하기 위하여 입력 음성파형과 출력 음성파형을 비교하여 제안한 알고리즘의 효과를 확인한다.

• 말소리와 음성과학
• /
• 제1권3호
• /
• pp.95-102
• /
• 2009

In this paper, an efficient target speech detection algorithm is proposed for the performance improvement of multi-input speech enhancement. Using the normalized cross correlation value between two selected channels, the proposed algorithm estimates the probabilistic distribution function of the value from the pure noise interval. Then, log-likelihoods are calculated with the function and the normalized cross correlation value to detect the target speech interval precisely. The detection results are applied to the generalized sidelobe canceller-based algorithm. Experimental results show that the proposed algorithm significantly improves the speech recognition performance and the signal-to-noise ratios.

• 한국멀티미디어학회논문지
• /
• 제19권7호
• /
• pp.1127-1136
• /
• 2016

Speech enhancement algorithm plays an important role in numerous speech signal processing applications. Over the last few decades, many algorithms have been studied for speech enhancement. The algorithms are based on spectral subtraction, Wiener filter, and subspace method etc. They have good performance of speech enhancement, but the performance can be deteriorated in specific noises or low SNR environment. In this paper, a new speech enhancement algorithms are proposed based on adaptive noise canceller. And the proposed algorithm improved performance of adaptive noise cancelling using 2-D binary mask. From objective experimental index, it is confirmed that the proposed algorithm is useful and has better performance than recently proposed speech enhancement algorithms.

• ETRI Journal
• /
• 제38권2호
• /
• pp.366-375
• /
• 2016

This paper describes an algorithm to suppress composite noise in a two-microphone speech enhancement system for robust hands-free speech communication. The proposed algorithm has four stages. The first stage estimates the power spectral density of the residual stationary noise, which is based on the detection of nonstationary signal-dominant time-frequency bins (TFBs) at the generalized sidelobe canceller output. Second, speech-dominant TFBs are identified among the previously detected nonstationary signal-dominant TFBs, and power spectral densities of speech and residual nonstationary noise are estimated. In the final stage, the bin-wise output signal-to-noise ratio is obtained with these power estimates and a Wiener post-filter is constructed to attenuate the residual noise. Compared to the conventional beamforming and post-filter algorithms, the proposed speech enhancement algorithm shows significant performance improvement in terms of perceptual evaluation of speech quality.

• 음성과학
• /
• 제10권2호
• /
• pp.77-84
• /
• 2003

In this research, we seek the beginning of the speech and detect the stationary speech region using lip information. Performing running average of the estimated speech signal in the stationary region, we reduce the effect of musical noise which is inherent to the conventional MlMSE (Minimum Mean Square Error) speech enhancement algorithm. In addition to it, SFM (Spectral Flatness Measure) is incorporated to reduce the speech signal estimation error due to speaking habit and some lacking lip information. The proposed algorithm with Wiener filtering shows the superior performance to the conventional methods according to MOS (Mean Opinion Score) test.

• 말소리와 음성과학
• /
• 제3권3호
• /
• pp.91-98
• /
• 2011

In this paper, a multi-channel speech enhancement method using the linearly constrained minimum variance (LCMV) algorithm and a variable learning rate control is proposed. To control the learning rate for adaptive filters of the LCMV algorithm, the direction of arrival (DOA) is measured for each short-time input signal and the likelihood function of the target speech presence is estimated to control the filter learning rate. Using the likelihood measure, the learning rate is increased during the pure noise interval and decreased during the target speech interval. To optimize the parameter of the mapping function between the likelihood value and the corresponding learning rate, an exhaustive search is performed using the Bark's scale distortion (BSD) as the performance index. Experimental results show that the proposed algorithm outperforms the conventional LCMV with fixed learning rate in the BSD by around 1.5 dB.

• 전자공학회논문지
• /
• 제54권5호
• /
• pp.85-89
• /
• 2017

본 논문에서는 향상된 연산 능력을 가진 하드웨어와 알고리즘의 혼합을 통하여 음성 향상을 위한 정확한 음성 검출기 구현을 목적으로 하였다. 음성은 음소의 나열로 구성되어있으며 음성 모델을 세우는데 적합한 방법은 이전의 정보를 이용하는 순환 신경망 (recurrent neural network, RNN)을 사용하는 것이다. 실제 존재하는 모든 잡음에 대하여 학습한 모델을 제시하는 것은 사실상 불가능 하므로 이를 극복하고자 음소기반 학습을 진행하였다. 학습의 결과로 세워진 모델을 기반으로 새로운 음성 신호에서 음성을 검출하고 그 결과를 이용하여 음성 향상을 진행하였다. 순환 신경망과 음소기반 학습은 프레임 별 높은 상관성을 가진 음성 신호에서 좋은 성능을 얻을 수 있었으며 음성 검출기의 성능을 검증하기 위하여 라벨 데이터와 음성 검출결과를 비교하고 다양한 잡음 환경에서 객관적 음질 평가를 진행하여 기존의 음성 향상 알고리즘과 비교하였다.