Development of a speech enhancement dual-microphone noise reduction system utilizing TMS320C6713

Autor: Zeng-fong Chen, 陳政鋒
Rok vydání: 2015
Druh dokumentu: 學位論文 ; thesis
Popis: 103
The purpose of this research was to add a speech enhancement process that could further improve speech intelligibility and the performance of automatic scene classification and auto-matching noise reduction system after the application of the adaptive directional microphone strategy. The speech enhancement system is divided into two parts, one is the noise-estimation strategy and another the speech-estimation function. Noise-estimation algorithms used in the research are: Minimum Statistics (MS), Minima-Controlled Recursive Averaging (MCRA), Improved Minima-Controlled Recursive Averaging (IMCRA), Minima-Controlled Recursive Averaging-Loizou (MCRA-L), Constrained Variance Spectral Smoothing (CVS), Forward-Backward MCRA(MCRA-FB); Speech-estimation function: Maximum-Likelihood (ML), Log-Spectral Amplitude (LSA), Maximum A Posteriori Amplitude (MAPA), Wiener-type, Wiener Filter. In this research, The MATLAB (The MathWorks, Natick, Massachusetts, USA) software was first used to simulate the speech enhancement system. The simulation was mainly to evaluate the speech quality of the signal after speech enhancement process with different signal-to-noise ratio (SNR) of the input speech noise signal, and then to select the best combination of the speech enhancement system. Finally, the selected speech enhancement system was implemented with automatic scene classification and auto-matching noise reduction system in TMS320C6713 DSP Starter Kit (Texas Instruments, Dallas, Texas, USA), and compared with the output signal in the original noise reduction system. To show the performance of the selected speech enhancement system, the objective perceptual evaluation of speech quality (PESQ) approach and the subjective speech reception threshold (SRT) were further used to evaluate the quality of speech with the SNR range between 30dB to -30dB. In the objective evaluation, the simulated results showed that the PESQ score was increased by 0.45 when the speech enhancement CVS with MAPA was used for the input signal with 30dB SNR and by 0.65 for 10 dB SNR. For the hardware implementation, only the speech enhancement MCRA with MAPA was used for real-time processing. The experimental results indicated that speech enhancement system could decrease the speech quality by 0.36 for the input signal with 30dB SNR. When the SNR was below 10dB, the automatic scene classification system would automatically select the function of microphone noise reduction strategy. With the speech enhancement system, our overall hardware implementation could effectively reduce speech distortion and improve speech quality. The PESQ score was increased by 0.27 for the input signal with 0 dB SNR. The SRT from five normal hearing subjects (between 23 to 26 years old) in different noise conditions were measured with the HINT Pro system (Bio-logic, Chicago, IL, USA) for subjective evaluation. Our experimental results showed that speech enhancement could not improve the SRT of the subjects, but become worse than original system. The average SRT of the subjects was increased by 8.54dB because the volume of the signal processed by the speech enhancement system became too small, even though the objective speech quality was improved. The above-mentioned experimental results suggested that the speech enhancement system could provide better speech quality in high SNR when the system used shorter frame length despite of some distortion in low SNR. Nevertheless, the speech enhancement system was able to greatly improve speech intelligibility when the system used longer frame length. If the amplifier stage was included in the system, the whole system could achieve the same performance as that of the objective evaluation.
Databáze: Networked Digital Library of Theses & Dissertations
Externí odkaz: