Design of a speech coder utilizing speech recognition parameters for server-based wireless speech recognition

The existing standard speech coders have good speech communications quality; however, it is known that they degrade the performance of automatic speech recognition (ASR) systems. that are deployed for wireless communications as a server-based approach. The paper proposes a speech coder that utilizes speech recognition parameters for wireless speech recognition. To maintain the performance of ASR as in conventional ASR systems implemented in a client-based approach, the proposed speech coder first extracts Mel-frequency cepstral coefficients (MFCC) that are the typical recognition parameters, and then converts them into linear prediction coefficients for CELP-type speech coding. By transmitting MFCC directly to the decoder, an ASR system employing the proposed speech coder can provide even better performance than that using standard CELP-type speech coders.