Optimizing speech quality and intelligibility through noise-aware enhancement for coding application

Significant advancements in speech coding research have been made in recent years, yet effect of background noise remain as a major cause of performance degradation. Some form of noise suppression pre-processing seems necessary under real world noisy environments. However, a single enhancement solution does not perform well across diverse noise types and levels. In the current study, we present a noise-aware enhancement pre-processing framework designed to improve both speech quality and intelligibility for linear predictive coding (LPC)-based system. This approach integrates noise classification with adaptive enhancement, enabling the selection of suitable algorithms for different noise conditions prior to encoding. Our experiments on the NOIZEUS corpus indicate that LPC speech coding can benefit from noise-aware enhancement pre-processing framework in low signal-to-noise ratio (SNR) conditions. In comparison with baseline noisy speech and enhanced speech coding, we have found that the proposed coding system performs considerably better in terms of quality and intelligibility metrics. The source codes developed for this work can be accessed at https://github.com/dathu/BGN_TRY_RGP_Noise-Type-identification/tree/main .