Traditional parametric vocoders are constructed based on a discrete time-domain speech signal generation model, primarily consisting of three parts: an excitation source, an acoustic model, and a radiation model. For voiced sounds, a periodic pulse generator is used to produce the pitch frequency; for unvoiced sounds, the excitation signal is generated by a random noise generator. The signal first passes through a glottal model, then enters the vocal tract model via a voiced/unvoiced switching module, where the vocal tract parameters determine the timbre characteristics of the speech. Finally, the signal is processed by the radiation model to form the final speech waveform output.

 

 

At the encoding end, the algorithm performs Linear Predictive Analysis on the input speech to extract vocal tract spectral envelope parameters (such as LPC/LSF), while simultaneously estimating excitation-related parameters including the pitch period, gain, and band-split voiced/unvoiced features. By modeling the spectral structure of the residual signal, it achieves the separation and quantized representation of periodic and noise components. At the decoding end, a mixed excitation signal is reconstructed based on the transmitted parameters, and the speech waveform is restored through a synthesis filter.