Floreat’s VQ (Vector Quantized) Speech Compression Algorithm
(VQ Proprietary Speech Codec Software)

Floreat’s Vector Quantization (VQ) speech compression algorithm is designed to provide maximum fidelity with minimum compute power and bandwidth. The VQ algorithm has two major components. The first section reduces the dynamic range of the input speech signal by removing short term and long term redundancies. This reduction is done in the waveform domain, with the synthesized part used as the reference for determining the incremental "new" content. The second section maps the residual signal into a code book optimized for preserving the general spectral shape of the speech signal.

The major advantages of this approach over other low-bit-rate algorithms are that there is no need for any intensive complicated calculation of reflection coefficients. Also, the quantization noise in the output speech is hidden under the speech signal and there are no pitch tracking artifacts; the speech sounds "natural", with only minor increases of background hiss at lower bit-rates. The computational load is reduced significantly compared to a VSELP algorithm and only minor variations of the same algorithm provide bit rates of 8, 9.2 and 16 Kbps. The total delay through the analysis section is less than 20 milliseconds. The present algorithm is accomplished completely in the waveform domain and there is no spectral information being computed and there are no filter computations needed.

The VQ algorithm relies on the fact that successive speech samples are correlated. Ideally, only the incremental innovations (changes) need to be transmitted. The algorithm operates on 160 digitized speech samples (20 milliseconds), divides the speech samples into time segments of 5 milliseconds each, and uses predictive coding on each segment. With the VQ algorithm, the current segment is predicted as best as possible based on the past recreated segments and a difference signal is determined. The difference value is compared to the stored difference values in a lookup table or code book, and the address of the closest value is sent to the remote site along with the predicted gain and pitch values for each segment. In this fashion, four 5 ms speech segments can be reduced to a packet of 23 bytes or 184 bits (46 bits per sample segment). By transmitting 184 bits every 20 milliseconds, an effective serial data transmission rate of 9200 bps is accomplished. By varying the table sizes and the quantization levels for each segment, we also create bit rates of 8 Kbps and 16 Kbps.

Floreat’s VQ speech compression software is widely deployed by many large organizations and US government agencies for their proprietary specialized applications.

Floreat’s VQ proprietary speech codec software can be integrated with Floreat’s other fax, telephony, ITU-T speech compression, VoIP, FoIP, DSVD, imaging and video software for various applications. Floreat’s VQ audio codec software is supported on various DSPs and processors as well as offered in ANSI C.

Top