Segmented dimensionality reduction coding on frequency domain signal

This paper proposes schemes of compressing frequency domain signals using dimensionality reduction methods. Dimensionality reduction methods which work on a input matrix result in two dimension-reduced output matrices with usually high compression ratio since they not only allow us to represent the input matrix with smaller amount of data, but exploit intrinsic information of the original data. Frequency domain signals can be seen as a (number of frequency bands) × (number of total frames) input matrix of dimensionality reduction methods. However, in this case, real-time encoding is not possible and encoder-side delay is inevitable which amounts to the length of whole input signal. To minimize the delay this paper proposes a coding scheme which conducts multiple dimensionality reduction on segments of input data frames serially.