Abstract

In this paper it is considered the problem of reduction or reduction of the order p ≫ 1 of an autoregressive model (AR-model) of a speech signal by the criterion of minimum loss of useful information. The problem is formulated as an optimization problem in terms of discrete spectral modeling. It is indicated that the most acute problem in solving is the necessity to scale the AR-model parameters for the simulated signal at each step of iterative calculation process. To overcome this problem, it is proposed to use the measure of information divergence of signals in the frequency domain with the property of scale invariance as the goal functional. On its basis, a new method of the AR-model reduction is developed where the scaling operation exceeds the limits of the iterative optimization procedure. The effectiveness of the proposed method is substantiated theoretically and researched experimentally. It is shown that the main component of the achieved effect is the gain in accuracy of the reduced AR-model in the Kullback–Leibler information metric. The results obtained are addressed to researchers and developers of systems and technologies for digital speech transmission over low-speed communication channels.

中文翻译：

---

低速通信信道上语音传输系统语音信号自回归模型的简化方法

摘要

在本文中，它被认为是减少或减少阶p的问题≫ 1 个语音信号的自回归模型（AR 模型），根据有用信息的最小损失标准。该问题被表述为离散谱建模方面的优化问题。指出解决中最尖锐的问题是在迭代计算过程的每一步都需要对模拟信号的AR模型参数进行缩放。为了克服这个问题，提出了以尺度不变性为目标泛函的频域信号信息散度度量。在此基础上，开发了一种新的 AR 模型缩减方法，其中缩放操作超过了迭代优化过程的限制。所提方法的有效性在理论上得到了证实，并在实验上进行了研究。结果表明，所取得的效果的主要组成部分是在 Kullback-Leibler 信息度量中减少 AR 模型的准确性增益。获得的结果是针对在低速通信信道上进行数字语音传输的系统和技术的研究人员和开发人员。