• The Journal of the Acoustical Society of Korea
• /
• 제29권2E호
• /
• pp.86-99
• /
• 2010

This paper analyzes the performance of various single channel speech enhancement algorithms when they are applied to automatic speech recognition (ASR) systems as a preprocessor. The functional modules of speech enhancement systems are first divided into four major modules such as a gain estimator, a noise power spectrum estimator, a priori signal to noise ratio (SNR) estimator, and a speech absence probability (SAP) estimator. We investigate the relationship between speech recognition accuracy and the roles of each module. Simulation results show that the Wiener filter outperforms other gain functions such as minimum mean square error-short time spectral amplitude (MMSE-STSA) and minimum mean square error-log spectral amplitude (MMSE-LSA) estimators when a perfect noise estimator is applied. When the performance of the noise estimator degrades, however, MMSE methods including the decision directed module to estimate a priori SNR and the SAP estimation module helps to improve the performance of the enhancement algorithm for speech recognition systems.

• Journal of information and communication convergence engineering
• /
• 제10권2호
• /
• pp.162-167
• /
• 2012

This paper proposes a speech processing system based on a model of the human auditory system and a noise reduction neural network with fast Fourier transform (FFT) amplitude and phase spectrums for noise reduction under background noise environments. The proposed system reduces noise signals by using the proposed neural network based on FFT amplitude spectrums and phase spectrums, then implements auditory processing frame by frame after detecting voiced and transitional sections for each frame. The results of the proposed system are compared with the results of a conventional spectral subtraction method and minimum mean-square error log-spectral amplitude estimator at different noise levels. The effectiveness of the proposed system is experimentally confirmed based on measuring the signal-to-noise ratio (SNR). In this experiment, the maximal improvement in the output SNR values with the proposed method is approximately 11.5 dB better for car noise, and 11.0 dB better for street noise, when compared with a conventional spectral subtraction method.

• 융합정보논문지
• /
• 제7권6호
• /
• pp.159-166
• /
• 2017

본 논문에서는 방송 뉴스에서 화자 변화 검증 성능 향상을 위해서 입력소음음성 향상과 SNR(Signal to Noise Ratio)기반 가중 함수 $w_m$를 적용한 KL 거리 $D_s$를 실험하였다. GMM-UBM(Gaussian Mixture Model-Universal Background Model) 기반 KL(Kullback Leibler) 거리 D를 이용한 화자 변화 검증 시스템(실험 0)을 기본 시스템으로 한다. 실험 1은 실험 0의 입력소음음성 향상을 위해 MMSE Log-STSA(Minimum Mean Square Error Log-Spectral Amplitude Estimator)를 적용하였다. 실험 2는 실험 1의 기존 KL거리 D 대신에 $D_s$를 적용하였다. 실험 데이터베이스는 다양한 소음을 반영하기 위해 스포츠 뉴스와 실외 인터뷰를 중심으로 구축하였다. 실험은 화자 변화 정보의 누락을 막기 위해 MDR(Missed Detection Rate) 0%를 기준으로 하였다. 실험 0은 FAR(False Alarm Rate) 71.5%의 성능을 보였다. 실험 1은 FAR 67.3%로 실험0에 비해 4.2% 향상되었고, 실험 2는 FAR 60.7%로 10.8% 향상되었다.