Evaluation of Speech Information Compression Coding System in Noisy Environments

语音信息压缩编码是一种适用于诸如调制解调器站和国际电话等相对较窄的频率范围的语音频带的有效语音传输方法。诸如ADPCM、ATC和LPC等语音编码系统最近已投入实际使用。然而，这些系统受到背景噪声的严重影响。研究的目的是为在噪声环境中建立高质量的语音传输系统奠定基础。在语音产生过程和噪声产生过程的建模、自适应卡尔曼滤波和语音合成系统等方面进行了改进，取得了以下成果。提出了一种有色噪声污染语音的语音/静音判决方法。该方法基于语音和噪声的短时自相关函数的性质。在检测到语音活动开始后，还提出了自适应谱估计方法来同时估计噪声和语音频谱。将所提出的方法应用于空调和计算机产生的室内噪声中的语音信号。从合成语音和自然语音的谱估计精度的角度验证了该方法的有效性。提出了一种基于自适应卡尔曼滤波的语音信号与加性噪声分离方法，对语音产生过程的状态空间表示中的系统参数进行自适应估计。通过对合成语音和自然语音的实验，证实了自适应卡尔曼滤波对浊音和清音都有一定的降噪效果，但对浊音的降噪效果大于对清音的降噪效果。通过对重建语音的主客观评价测试，验证了该系统的有效性，基本实现了本研究的初步目标。

A speech information compression coding is an effective speech transimission method for the speech band limited to relatively narrow frequency range such as modile stations and international telephone. Speech coding systems such as ADPCM, ATC and LPC have lately been out into practical use. However these systems are highly degraded by background noise. The purpose of the investigation is to establish the basis of a high quality speech transmission system in the noisy environments. The following results were obtained by improving the previous our works in the modeling of speech production process and noise generation process, adaptive Kalman filtering and speech synthesis system.1. A method for speech/silence decision of speech corrupted by colored noise was proposed. This method is based on the properties of the short-time autocorrelation function of both speech and noise. After detecting the onset of speech activity, the adaptive spectrum estimation method was also developed to estimate both noise and speech spectrum.2. The proposed methods were used for the speech signal uttered in the room noise generated by air conditioner and computer. The validity of the methods were confirmed from the point of view of spectral estimation accuracy for both synthetic and natural speech.3. A method based on the adaptive Kalman filter was proposed to separate the speech signal from additive noise, in which the system parameters in the state-space representation of the speech production process were estimated adaptively.4. From the experiments for both synthetic and natural speech, it was confirmed that the adaptive Kalman filtering has the effect of noise reduction for both voiced and unvoiced speech, but the effect is larger for voiced speech than for unvoiced speech. The validity of the proposed system was also demonstrated by objective and subjective evaluation tests of the reconstructed speech.As ahown above, the initial goal of this study was almost accomplished.