グルーンライナー,シンガーハラルド
Scalar Quantization of Cepstral
Parameters for Low Bandwidth
Client-Server Speech Recognition
Systems
Abstract:We describe a simple but highly efficient approach to solve the bandwidth problem for a client-
server architecture speech recognition system. Data transmission from client to server follows an
approach proposed by SRI, i.e. scalar quantization of cepstral parameters is used for compression.
Recognition results on ATR's Travel Arrangement task show that recognition error only increases
by about 0.5 % at a bandwidth of less than 6 Kb/s. This report is accessible for ITL members
via /home/singer/tex/TR-IT-Compress/.