ROBUST LOW BIT RATE ANALYSIS-BY-SYNTHESIS PREDICTIVE SPEECH CODING

Redwan  A. SALAMI

High quality analysis‑by‑synthesis predictive coding of speech signals at bit rates below 9.6 kb/s, and the application of error‑correcting coding to the digitally encoded speech to improve the transmission performance over mobile radio channels are investigated. Analysis‑by‑synthesis predictive coding is a new class of speech coding able to produce high‑quality speech in the bit rate range from 4.8 to 16 kb/s. However, at bit rates below 8 kb/s, the encoding algorithms become computationally demanding due the vector quantization of the excitation frames. As digital mobile radio telephony is one of the most important application of this class of speech coding, it is very important to improve the coder robustness, and to efficiently utilize error‑control coding to guarantee high quality speech reception over the hostile mobile radio channel. The work in this thesis has concentrated on reducing the complexity of the encoding algorithms while maintaining high speech quality below 8 kb/s, and error‑protecting the digitally encoded speech in an efficient manner with the lowest amount of redundancy.

Multi-pulses excited and regular‑pulse excited LPC coders have been studied in detail, and high quality speech at bit rates around 9 kb/s has been obtained. Code‑excited linear prediction (CELP) coding has been investigated, and several methods for reducing the coder complexity have been discussed. Communication quality to high quality speech was produced in the bit rate range from 4.8 to 8kb/s. A novel approach, which dramatically reduced the CELP coding complexity, has been proposed. The binary pulse excitation (BPE) approach was able to deliver the same speech quality of CELP coders utilizing a very efficient encoding algorithm in which the exhaustive search of the excitation codebook is completely ruled out. The BPE also significantly reduced the storage requirements of the excitation codebooks , where no storage was needed in case of nontransformed excitation pulses. Transforming the excitation pulses enabled the use of the equivalent of huge excitation codebooks, which significantly improved the speech quality.

As a proposal for the half‑rate GSM coder specifications, a high quality 7.4 kb/s BPE coder is described . The robustness of the speech coder against channel errors was given a high priority while selecting the coder parameters. Error‑correcting coding was embedded in the speech coder aiming at the overall bit rate of 11.4 kb/s. The sensitivity of the speech carrying bits was studied, and the error‑control bits were carefully allocated so that more important speech bits were efficiently protected. Both convolutional codes and Reed‑Solomon (RS) codes were studied, and the system was designed such that the same channel coder could provide different levels of protection. Puncturing was used in case of convolutional codes to achieve this purpose. In case of RS codes, a novel approach has been proposed in which different shortened versions of a primitive RS code enable the variation in the protection levels. The proposed RS coding scheme exhibited better performance than the convolutional coding scheme and it is simpler to implement. About 14 dB channel SNR is sufficient to maintain high quality speech reception over mobile radio channels.