Low-Bit Rate Speech Encoders Based on Line-Spectrum Frequencies (LSFs)

Abstract

Abstract : After nearly a dozen years of research and development, the narrowband linear predictive coder (LPC) operating at 2400 bits per second (b/s) has become a practical means to transmit speech at less than 5% of the bit rate of the original digitized speech. The 2400-b/s LPC is expected to be deployed extensively on various military platforms. Recently, however, there has been a growing interest in very-low-data-rate (VLDR) (800 b/s or less) voice communication for certain specialized military communications. Likewise, there is a demand for a voice processor operating at 4800 b/s that outperforms significantly the current 2400-b/s LPC is for operation in less than favorable environments. We present in this report a means for implementing 800- and 4800- b/s voice processors. Both processors use a similar speech synthesis filter in which control parameters are line-spectrum frequencies (LSFs). The LSFs are frequency-domain parameters transformed directly from the prediction coefficients used in the conventional-narrowband LPC. The use of frequency- domain parameters is highly significant because they allow filter coefficient quantization in accordance with properties of auditory perception (i.e., coarser representation of higher frequency components of the speech spectrum). The excitation signal at the 800-b/s rate is similar to that used in the conventional-narrowband LPC. On the other hand, the excitation signal at 4800 b/s is the baseband residual signal. At 800 b/s, speech intelligibility as measured by the diagnostic rhyme test (DRT) is 87 for three male speakers; this is about 1.4 points lower than that of the 2400-b/s LPC.