线性预测技术已广泛用于语言信号处理,特别是用于设计低比特率的声码器。但是传统的定帧长分析方法不能很好适应语言的非平稳过程,同时由此得到的语言参数(例如预测误差)用于提取音调也容易出错,因而影响了合成语言的质量。为此我们进行了三方面的改进:(1)用自适应梯型算法代替现有定帧分析的线性预测方法,以便得到更准确的声道参数;(2)有限的变帧抽样语言参数,改善了合成语音过渡区的性能;(3)改进Gold-Rabiner的基音提取技术,使音调提取方法更简单可靠。传送数据率为2400和1200比特每秒。在计算机上模拟结果表明,2.4kb/s的方案所合成的语言较为自然易懂,且不难分辨熟人口音,l.2kb/s方案的合成语言也未严重降级。采用缓存器后,两种方案均可在固定数据率信道上传输。 Linear prediction techniques have been widely used for speech signal processing, especially for the design of low bit rate vocoder. However, the traditional frame length analysis method can not well adapt to the language non-stationary process, and the resulting language parameters (such as prediction errors) are also error-prone for extracting tones, thus affecting the quality of the synthesized language. To this end, we make three improvements: (1) replace the linear prediction method of the existing frame analysis with adaptive ladder algorithm in order to obtain more accurate channel parameters; (2) limited variable sampling variable sampling language parameters, Improve the performance of the synthesized voice transition region; (3) improve the Gold-Rabiner’s pitch extraction technology to make the pitch extraction method more simple and reliable. The transmission data rate is 2400 and 1200 bits per second. Computer simulation results show that 2.4kb / s program synthesized by the language is more natural and easy to understand, and not difficult to distinguish acquaintances accent, l.2kb / s program synthesis language nor serious degradation. After adopting the buffer, two kinds of schemes can all transmit on the fixed data rate channel.