Colloquium - Details

Master-Vortrag: Development of a backward adaptive linear prediction scheme for the SCELP low delay audio codec

Srikanth Korse
24. Februar 2014
11:00 Uhr
Hörsaal 4G IKS

The SCELP low delay audio codec was recently proposed by IND to provide a novel candidate for digital wireless microphones in live concerts. It employs forward adaptive warped linear prediction in combination with spherical vector quantization and yields an algorithmic delay below 10 ms at a data rate of approximately 50 kBits/sec given a sample rate of 22 kHz. The SCELP codec requires buffering of the input signal for block adaptive LP analysis as well as the transmission of the LP coefficients with an average data rate of 5-6 kBit/sec. In order to achieve an even lower algorithmic delay and a lower overall bit rate, in this thesis, the use of backward adaptive linear prediction was investigated as an alternative to the block based approach as used in the proposed SCELP codec. For this purpose, sequential Least Mean Square (LMS) and sequential Gradient Adaptive Lattice (GAL) approaches were realized and evaluated for backward linear prediction analysis in combination with scalar quantization in the first part. In particular, these two approaches were extended to allow noise shaping techniques to improve the overall perceived audio quality. In the second part, the SCELP codec was extended to work based on backward adaptive linear prediction.
As a result, the SCELP could be modified to operate in backward adaptive manner to achieve an algorithmic delay of just one vector size (11 samples equivalent to 0.5 ms audio data) at a data rate of 44 kBit/sec with minimal or no perceivable quality loss compared to the original SCELP codec.

back