Low bit rate speech compression for playback in speech recognition systems

Abstract

In this paper we describe a novel, low complexity, low bit rate speech compression and decompression methods for usage in systems where automatic speech recognition is performed. The coding scheme, referred to as the Recognition Compatible Voice Coder (RECOVC), is based on encoding the mei-frequency cepstral coefficients (MFCC), commonly used in large vocabulary continuous speech recognition systems, and the pitch period. The decoder reproduces natural sounding, good quality, intelligible speech for playback purposes. Implementation of a RECOVC scheme in a speech recognition system may simplify the playback procedure by reconstructing speech from feature vectors already extracted and used for recognition. Reduction in storage space or transmission bandwidth may be achieved in distributed speech recognition systems, by eliminating the need to store or transmit two separate bit streams, one for recognition and the other for playback.