Sangbae Jeong

Robust GSC-based speech enhancement for human machine interface

This paper is concerned with generalized sidelobe canceller (GSC)-based speech enhancement. The speech enhancement is performed in the condition that the arrival direction of the target speech source is given by the sound source localization module. To compensate for the channel-mismatch of the input signals caused by the path differences, fixed and adaptive channel compensators are adopted before noise reduction. Then, the GSC-based speech enhancement is performed for the channel-mismatch compensated input signals. To improve the noise reduction performance, the probabilistic adaptation mode controller is introduced to the GSC. Our experimental results in home-robot environments show that the proposed noise reduction significantly improves the speech recognition performance in real noisy environments.

Adaptive microphone array processing for high-performance speech recognition in car environment

In this paper, a novel speech enhancement algorithm is proposed. The algorithm controls the amount of noise reduction according to information about speech absence or presence. Compared to conventional linearly constraint minimum variance beamforming, the proposed noise reduction achieved remarkable improvement in speech recognition rates.

Dual-Microphone Noise Reduction in Car Environments With Determinant Analysis of Input Correlation Matrix

In this paper, a novel noise reduction technique is proposed to improve the speech interface performance in car environments. The proposed noise reduction method with dual microphones is primarily based on the determinant analysis of the input correlation matrix. Through the analysis, a robust feature for speech activity detection and signal-to-noise ratio (SNR) estimation is derived. Using the feature, the SNR of each time-frequency component is estimated, and an enhanced speech signal is obtained through Wiener filtering. To evaluate the proposed noise reduction technique, we constructed a database in a real car environment and comparatively analyzed the performances of noise reduction methods. The results show that meaningful SNR and perceptual speech quality improvements with less signal distortion are achieved compared with the other competing methods.

Wiener filter-based echo suppression and beamforming for intelligent TV interface

Speech recognition for intelligent TVs is not easy mainly because of the TV sound itself. Input signals for automatic speech recognition systems have a low SNR condition due to the sounds from the TV acoustic speakers near to the microphone array installed on a TV. In addition, spoken commands for TV control are usually given at a considerably far distance. This tends to cause reverberated command inputs easily corrupted by other environmental noises. To achieve successful speech recognition with the harsh inputs, a powerful noise reduction algorithm is proposed. It is a combined solution cascading Wiener filter-based acoustic echo suppression (AES) and adaptive beamforming. To obtain noise power for AES, reference noises are estimated by utilizing the input signals to the TV speakers. For evaluation, output SNRs and speech recognition rates were measured under various noisy conditions and the results of the proposed system showed significant improvements, especially for low SNR.

Performance improvement of signal classifiers for speech/audio coding by higher-order statistics

In this paper, a new decision rule for the frame-by-frame classification of speech and music signals is proposed. Based on the signal kurtosis and skewness, a binary decision tree is constructed by the classification and regression tree algorithm. By the speech quality tests in G.718 speech/audio coder, the superiority of the proposed algorithm is confirmed.

An efficient approach using HOS-based parameters in the LPC residual domain to classify breathy and rough voices

Abstract Although a considerable number of studies have been focused on the analysis of pathological voices using conventional parameters such as jitter, shimmer, and signal-to-noise ratio (SNR), these parameters have been found to be sensitive to variations in pitch extraction algorithm and cannot analyze severely disordered voice signals which exhibit irregular or aperiodic waveforms. In this paper, higher-order statistics (HOSs) analysis, which is independent of pitch period, is derived from linear predictive coding (LPC) residuals to describe breathy and rough voices. Recordings of a sustained /a/ from 23 individuals with breathy voices and 30 individuals with rough voices were collected from the disordered voice database distributed by the Japanese Society of Logopedics and Phoniatrics. We extracted conventional parameters as well as the HOS-based parameters such as the normalized skewness and the normalized kurtosis. On the other hand, we calculated HOS-based parameters from the LPC residual domain. The results showed that the HOS-based parameters and the HOS-based parameters estimated from the LPC residual are different for rough and breathy voices. Conventional parameters were not distinctive for these voices. Classification and regression tree (CART) was used to combine multiple parameters and to classify breathy and rough voices. Using the HOS-based parameters, the CART achieved an accuracy of 85.0% with the optimal decision tree generated by means of the normalized skewness and kurtosis. When the HOS-based parameters using LPC residual were used, the optimal decision tree was 88.7% accurate and the variances of the normalized skewness and kurtosis were included.

Performance improvement of speaker recognition by MCE-based score fusion

This paper proposes an improved feature extraction in glottal flow signals and minimum classification error-based fusion of multiple feature scores to improve speaker recognition accuracy. Experimental results show that the proposed score fusion method with the improved feature extraction in glottal flow signals outperforms conventional speaker recognition methods.

Cepstrum-Based Bandwidth Extension for Super-Wideband Coders

This letter proposes a bandwidth extension (BWE) method using the cepstral envelope coding and duplication of quantized wideband (WB) signals by means of analysis -by-synthesis (AbS) for super-wideband (SWB) coders. In the proposed method, a high frequency band is generated by utilizing the quantized cepstral coefficients extracted from the envelope and the quantized modified discrete cosine transform (MDCT) shape of the wideband signal. The proposed method is compared with the latest G.718 SWB codec and the experimental results show that the proposed method outperforms the baseline codec both in subjective listening tests and objective performance measures.

Frame error-robust MDCT bit reduction for G.729.1 by inter-frame correlation

In this paper, a new algorithm for encoding spectral envelope in G.729.1 for VoIP is proposed. In order to reduce allocation bits, sub-band correlation between adjacent frames is utilized. By the proposed allocation algorithm, enhanced quality of sounds can be obtained even in frame error conditions.

Multi-channel noise reduction with beamforming and masking-based Wiener filtering for human-robot interface

In this paper, an efficient noise reduction algorithm is proposed for robust speech recognition. For the nonstationary noise reduction, frequency-domain beamforming-based speech enhancement is performed and masking-based Wiener filter is applied to the beamforming output. To design the masking-based Wiener filter, the spectrum of beamforming output is classified into noise spectrum and speech spectrum at each spectral bin by the inter-channel time delay between two reference inputs. Hamming windowing for the speech spectrum and noise spectrum is separately performed to smooth each spectrum. Then, the Wiener filtering is applied to the beamforming output. The performance of the proposed algorithm significantly improves the speech recognition accuracies and the signal-to-noise ratios.