REPRESENTASI EFISIEN DARI SPEKTRUM SINYAL UCAPAN-BERBAHASA INDONESIA BERBASIS LINE SPECTRUM PAIRS (LSP)

• Main Author: Marpanaji , Eko
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/2959
• Institution: Institut Teknologi Bandung

<b>ABSTRACT:</b><br> <br /> Speech signal processing based on Linear Predictive Coding (LPC) is widely used in many applications, using LPC coefficients as spectral representation of speech signal. For low bit rate application, quantization of LPC coefficients is a serious problem. Direct quantization of predictor coefficients is usually avoided because predictor coefficients are very sensitive to quantization errors and need few bits. <br /> <br /> In order to reduce quantization errors, another spectral representation called Line Spectrum Pairs (LSP) method can be used instead of LPC coefficient. <br /> With this method, scalar quantization can be applied for each LSP frequency. The object of this research is how to define quantization levels for each LSP frequency using statistical approach, such that the quantization levels can be used for spectral representations of speech signal based on Indonesian Language. <br /> <br /> Using more than 22,000 frames with 20 speakers (males and females) for investigating statistical distributions during training process, the table of quantization levels can be defined. As a result, spectral representations using LSP just need 34 bits with means of spectrum distortion less than 1 dB (0,77619 dB). <br /> <br /> The result of this research is that the table of scalar quantization of LSP frequencies can be used for spectral representation of speech signal based on Indonesian Language in low bit rate (2,400 bps) applications. In this research, the application of speech signal processing is focused on digital communications.