A low-complexity waveform interpolation coder

Abstract

A B S T R A C TA recent independent survey found a 2.4 kb/s waveform-interpolation (WI) algorithm to perform better than other state-of-the-art coders. However, this coder had a very high level of computational complexity. The introduction of various techniques, including a time-varying waveform sampling rate and a cubic B-spline waveform representation, has reduced the computational complexity by an order of magnitude. The new implementation allows full-duplex reaktime operation on a single DSP device and on an average workstation (the latter using nonoptimized, compiled C source code). The new coder also contains a number of new features which improve the quality of the reconstructed speech signal. I N T R O D U C T I O NIn the last ten years, the performance of 2.4 kb/s speech coders has increased significantly, resulting in new applications and a need to replace the existing $31015 standard [l]. The waveform interpolation (WI) coding paradigm [2] was recently found to provide state-of-the-art performance at 2.4 kb/s [l]. The coder performed very well in terms of quality, robustness against channel errors and background noise, and speaker recognizability [3]. In all of these tests, the 2.4 kb/s WI coder was found to be statistically equivalent to, or better than, the 4.8 kb/s FS10lG standard [4].To make the WI coder useful for practical applications, it must be implementable on a single DSP. We describe procedures which reduce the computational complexity to a level that permits this. In addition, we describe procedures to improve the reconstructed speech quality.The outline of this paper is as follows: section 2 reviews the basic structure of WI coders; sections 3 and 4 describe our improved WI implementation. Section 5 provides a complexity overview and section G provides conclusions.2. WI P R I N C I P L E S evolving waveform. A two-dimensional signal, g ( t , d), is defined, which displays the representative shape of the time waveform along the 4 axis. The waveform is repeated along the 4 axis with a normalized period of 27r. For voiced speech, the waveform evolves relatively slowly and for unvoiced speech the waveform evolves rapidly. Figure 1 provides a block diagram of the current 2.4 kb/s WI coder, including the sampling rates of the outputs of the blocks. The first step in the coding process is the linearprediction (LP) analysis, which is performed at a rate of Sf (the frame rate; 40 Hz in our implementation). The LP In WI codinp [ 2 ] , the speech sipnal i s represented by an alignment power h determination scale interpolate I synthesize inter olate 7 Figure 1. Basicoperationsin the 2.4 kb/s WI. The samplingrate of the output of each processing unit is shown on the right. For clarity, spectral and power quantization have not been depicted. parameters are quantized using a 30-bit split VQ with linear mapping of block codes [5] to obtain robustness against channel errors. The LP parameters are conventionally interpolated to a subframe update rate of S,f (160 Hz). Filtering the speech signal, s ( t ) ...

Enhanced waveform interpolative coding at low bit-rate

“…The interpolated SEW vectors are given by (7) Assuming and the LPC coefficients are given, the encoder's task is to find the quantized vector such that the accumulated weighted distortion between original and reconstructed waveform sequences, denoted by , is minimized. Since the effect of the linear interpolation LPF is taken into account in the proposed scheme, a true interpolated waveform (synthesis) is incorporated in the analysis process, unlike the conventional open-loop WI coders [10]- [18] in which only one waveform, namely , is used for the quantization. Consider the accumulated weighted distortion, , between the input SEW FCs vectors, , and the quantized and interpolated vectors, , given by (8) where number of waveforms per frame; number of look-ahead waveforms; diagonal matrix whose elements, , are the spectral values of the combined spectral-weighting and synthesis filters at the th harmonic given by (9) where pitch period; number of harmonics; gain; and input and the quantized LPC polynomials, respectively.…”

“…Most WI coders [10]- [18] use nonideal low-pass filters for downsampling and upsampling of the SEW. These filters introduce aliasing and mirroring distortion, even when no quantization is applied.…”

“…The interpolated pitch values are used for pitch cycle waveform extraction, which is performed at a regular rate (every 2 ms). The rate must be higher then the maximal pitch frequency in order to prevent aliasing along the time axis [14], [18]. The extracted waveforms are then power normalized, and sequentially aligned, to form a discrete-time CW, which is represented by a Fourier series (FS).…”

Predição linear harmônica para processamento espectral e temporal de sinais de voz.

“…O procedimento padrão de extração da forma de onda aplica amostragem uniforme [15] mas a extração crítica de ciclos de pitch tem sido usada para reduzir a complexidade da codificação [16] bem como para permitir a reconstrução perfeita da forma de onda [17].…”

Use of the pitch synchronous wavelet transform as a new decomposition method for WI

“…Currently in the WI coder, a trained codebook is used for the quantisation of the SEW magnitude information [6] which is transmitted once per frame. The REW magnitude spectrum is quantised using Chebyshev polynomial techniques and is transmitted multiple times per frame.…”