Techniques for low bit rate speech coding using long analysis frames

Abstract

Ericsson Radio Systems AB, S-164 80 Stockholm SWEDEN * Erisoft AB, S-951 10 Lule% ABSTRACTThe bit rate of CELP coders can be significantly reduced by increasing the length of the analysis frame and the subframes. When frame lengths of 30-40 ms and above are used, new problems that reduce speech quality do become prominent. The paper discusses the origin of some of these problems, and suggests techniques that can be applied in order to improve quality and reduce computational complexity for long frames and subframes. Promising results have been obtained for a new interpolation constrained time variable LPC algorithm and, in particular, for a new double mode LTP structure. Experimental results on delayed decision tree coding and low complexity innovation coding are also presented. 11-6040-7803-0946-4/93 $3.00 0 1993 IEEE

Improved spectral tracking using interpolated linear prediction parameters

“…An LP analysis method which is based on maximizing the prediction gain when using the interpolated parameters to update the prediction filter at each subframe is presented in [4]. Since the derivation of the algorithm incorporates the interpolated parameters, a better spectral fit is obtained for the intermediate subframes.…”

Improvements of background sound coding in linear predictive speech coders

“…The performance of each codec was evaluated in informal listening tests by skilled listeners. Experiment 1: The four codecs were run confi gured as described in [1]- [4]. Conclusions: 1) Swirling exists in all tested coders, and the distortion increases with reduced bitrate.…”

“…The cod, ecs were fi xed and fl oating point versions of the 7.95 kbps VSELP [1], [2], the 4 kbps Ericsson D-AMPS half raie candidate [3], [4], and a standard LPC vocoder [9], forced to un voiced mode. The test material included fl at and band pass fi ltered (300-3400 Hz) car and babble noise (40 s each), as well as fl at and bandpass filtered speech with the above disturbances added (10-20 dB signal to noise ratio).…”

“…Perceptually, the original disturbance is trans formed into a sound with close resemblance to fl owing water. Swirling also occurs in other linear predictive codecs than the VSELP, like in [3], [4]. Also the back ward adaptive LD-CELP codec [5] was reported to suf fer somewhat from the problem with swirling, at least at 8 kbps [6].…”