Speech data compression for embedded systems

The paper considers speech data compression for low-cost embedded systems such as voice-related toys or devices with interactive sound-responses. We use a PC to generate and compress 8-bit-speech-data that has various features such as human speech, symphony and animal songs; the compressed data are then transferred to a masked-ROM. An Intel 8051 embedded chip is employed to expand the compressed speech data stored in a masked-ROM with size ranging from 1-8 MBytes. No RAM (read/write memory) is available in the system (to reduce production cost). The ROM also stores the program that is responsible for expanding the compressed data. Since the compression is done off-line, we use simple vector-quantization techniques to perform a first stage lossy compression of the data; the lossy-compressed data are then treated by a second stage lossless compression. Because of the the lack of RAM in the expansion stage, an LZ compression technique cannot be applied. Instead, we use a runlength-Huffman encoding scheme coupled with a unary-code to perform the lossless compression. Also, because of the slow processing speed of the processor in the expansion stage, a length-limited Huffman code must be used. We develop a technique that uses splaying to achieve this goal