Multi-frame Quantization of LSF Parameters Using a Deep Autoencoder and Pyramid Vector Quantizer

This paper presents a multi-frame quantization of line spectral frequency (LSF) parameters using a deep autoencoder (DAE) and pyramid vector quantizer (PVQ). The object is to provide sophisticated LSF quantization for the ultra-low bit rate speech coders with moderate delay. For the compression and de-correlation of multiple LSF frames, a DAE possessing linear coder-layer units with Gaussian noise is used. The DAE demonstrates a high degree of modelling flexibility for multiple LSF frames. To quantize the coder-layer vector effectively, a PVQ is considered. Comparing the discrete cosine model (DCM), the DAE-based compression shows better modelling accuracy of multi-frame LSF parameters and possesses an advantage in that the coder-layer dimensions could be any value. The compressed coder-layer dimensions of the DAE govern the trade-off between the modelling distortion and the coder-layer quantization distortion. The experimental results show that the proposed algorithm with determined optimal coder-layer dimension outperforms the DCM-based multi-frame LSF quantization approach in terms of spectral distortion (SD) performance and robustness across different speech segments.