A polynomial interpolation-based scheme for reducing bandwidth in distributed speech recognition system

In this paper, we propose a low bit-rate compression scheme in distributed speech recognition (DSR) system based on polynomial interpolation. Dimensionality reduction of a set of successive Mel frequency cepstral coefficients (MFCCs) is achieved by performing polynomial least squares fitting. A conventional vector quantization (VQ) is applied to the polynomial coefficients to achieve more than 58% of bandwidth reduction as compared to ETSI advanced front-end (ETSI-AFE) encoder. Evaluation performance has been conducted on the Aurora-2 database in clean and multi-condition training modes. With respect to ETSI-AFE, the results obtained with the proposed encoder show no significant degradation in term of overall recognition accuracy.