手部動作辨識語音系統之研究

Speech Production System for Hank Motion Identification

江智勝(Chi-Sheng Jiang)

指導教授 ： 徐良育

中原大學/工學院/醫學工程研究所/碩士(2002年)

https://doi.org/10.6840/cycu200200540

摘要

在日常生活上，聽障人士僅能運用手語、唇語，或是利用紙筆作為溝通的工具。至今，針對聽障人士仍無一個有效的輔具來改善溝通上的不便。因此，建構一可攜式的手部動作轉語音系統，消除聽障人士溝通上的困難，為本研究主要的目的。

本研究在語音模組中先利用Labview圖控操作軟體中的錄音程式，以8kHz的採樣率、16bit的形式將語音訊號擷取出，再以多脈衝激勵線性預測編碼(MPLPC)壓縮編碼，再將編碼訊息與語音程式結合存放於語音合成晶片(SPDS105A)中；控制模組中利用七通道肌電訊號擷取系統由MSP430F148內建之A/D，擷取手前臂肌電訊號，並透過三種基本特徵值處理，正規化後經前向式類神經網路辨識，同時將結果輸出至語音模組。系統測試實驗中，定義了7種手部動作，對象為四男二女，平均年齡23±2歲，測試時使測試者坐著，以右手配戴主動電極腕帶，每種手部動作重複作十次與二十次。本研究分別Matlab及 MSP430F148進行辨識，當一半為訓練樣本，一半當為測試樣本時，Matlab方式的辨識結果為77.13﹪及87.85﹪，而MSP430F148的辨識結果為64.75﹪及70.28﹪，兩種方式結果顯示，訓練樣本愈多，辨識率愈高；另一方面，Matlab離線分析的辨識率都比MSP430F148辨識率高，顯示定點數的運算使類神經網路產生較大的誤差，造成辨識率降低；另一方面系統手部動作即時辨識，辨識率只有32.86﹪，歸納原因有前向式類神經網路無學習的能力、電極產生位移、人為因素、MSP430F148運算精確度。

整體而言，本系統已有初步的系統架構，但本系統手部動作辨識率仍不足，這其中有許多問題尚待解決，包含電極重新設計，在單晶片上實現倒傳遞神經網路，相信對於系統的辨識率能大大的提升。

關鍵字

多脈衝激勵線性預測編碼 ； 聽障 ； 類神經網路 ； 主動電極

並列摘要

The hearing impaired person can make use of sign language, lip reading or paper and pen as tools to communicate with other. Until now, there is still no effective tool to improve their communication ability. Thus, to develop a portable system to translate hand motions into speech is the goal of this study.

In the voice module, the speech is digitized using Labview voice recording program with a sampling rate of 8000Hz and 16-bits format. The digitized speech is then coded using multi-pulse linear predictive coding (MPLPC) and programmed into speech synthesis chip (SPDS105A). In control module, using seven-channel EMG acquisition system and the build-in analog-to-digital converter in the MSP430F148, the forearm EMG signals are acquired. EMG characteristics are obtained using three feature extraction methods and normalized before inputted into the forward neuron network for identification. In the same time, the result is outputted to the voice module. To evaluate the proposed system, seven hand motions are defined. Six subjects are recruited including four male and two female. The average age is 23±2. During the experiment, subject sits in a chair and wearing the active electrodes for EMG measurement in their right arm. Every designated hand motions are repeated for 10 or 20 times. Both Matlab method and MSP430F148 are used for classification. When half of the EMG signals are used for network training and the other half are used for testing, the results of classification using Matlab are 77.13﹪and 87.85﹪. On the other hand, the results of classification using MSP430F148 are 64.75﹪and 70.28﹪. The results indicates that the more the training samples are the better the classification. On the other hand, the results using Matlab methods are better than MSP430F148. This indicates that the fix-point computation in neural network cause error and reduce the rate of classification. Additionally, in the test of real-time classification, the classification rate is 32.86﹪. The reasons for this low percentage may include the lack of learning ability in the forward neuron network that is used in this study, the displacement of electrodes, the error caused by fix-point computation and other human factors.

Although the proposed system has the basic structure, the classification rate for hand motion is still too low. There are still a lot of questions to be resolved including redesigning the active electrode, realizing back-propagation neuron network in the micro-controller. It is believed that with all these improvements the identification rate can be increased.

並列關鍵字

multi-pulse linear predictive coding ； hearing impaired ； active electrode ； artificial neuron network

參考文獻

[12]田榮雯著, “以FPGA實現倒傳遞類神經網路並應用於肌電圖分類”, 中原大學碩士論

[19]陳建宇著, “多電極式手部動作辨識系統”, 中原大學碩士論文,2001年6月

[2]S. Pouremehdi, “Microcomputer-based tactile hearing prosthesis,Proceedings of Third Annual IEEE Symposium , pp.117-122,IEEE,1990

Proceedings of the First Joint BMES/EMBS Conference, pp.656,1999

[4]K. Katsutoshi, “A Discrimination System Using Neural Network for EMG-