This paper presents an approach, referred to as frequency domain interpolation (FDI), for achieving high-quality speech at low bit-rates (4 kb/s and below) within reasonable complexity and delay. FDI methods, like the prototype waveform interpolation (PWI) methods, derive a prototype waveform (PW) at regular intervals of time. But, unlike PWI, there is no separation into a slowly evolving waveform (SEW) and a rapidly evolving waveform (REW) component. Instead, the PW is encoded after gain normalization in magnitude-phase form. The magnitude is modeled as a sum of mean and deviation values in multiple frequency bands and this model is quantized using switched backward adaptive VQ techniques. The phase information is represented as a composite vector of PW correlations in multiple frequency bands and an overall voicing measure. This information is quantized using a VQ at the encoder. At the decoder, a phase model is employed that uses the received phase (and magnitude) information to reproduce PWs with the correct periodicity and evolutionary characteristics. Speech is synthesized by interpolating the reconstructed PWs after gain adjustment and filtering it using the short-term predictor and a postfilter. The design of a 4-kb/s and a 2.4-kb/s FDI codec are presented in this paper and their performance is characterized in terms of delay, complexity, and subjective voice quality. The results confirm that FDI techniques have the potential for delivering high-quality speech at low bit-rates in a cost-effective manner.

Robust vector quantization of LSF parameters at a low bit rate is essential for voice coders operating below 5 Kbps. A novel aspect of the proposed technique is the use of decorrelated residual LSF vectors from speech mode based backward prediction along with a multi-stage VQ design. Rates as low as 12 bits per 20 ms speech frame for the stationary voiced speech mode and 22 bits/frame for unvoiced and non-stationary voiced frames are shown to result in efficient quantization. In our classification scheme, spectrally stationary voiced frames constitute around 30% of active speech frames resulting in a minimum average bit rate of 19 bits/frame. The objective VQ performance is compared with cellular standard coders such as the IS-641 and IS-127. The proposed VQ has been integrated into a speech mode based 4.8 Kbps coder resulting in subjective performance close to that of the 7.4 Kbps IS-641 coder.