The source-filter paradigm, which is used in human speech development, would be used in this proposal to increase speech quality. Excitation signals in the cepstral domain are amplified using a sophisticated neural network (DNN). As a result, they compared the effects of normalization on two different types of goals. Using the cepstral excitation manipulation approach (CEM) instead of traditional signal processing-based cepstral excitation, we were able to reduce noise by 1.5 dB. Studying various combinations of data demonstrates that envelope and excitation enhancement are also present. When adopting a low signal-to-noise ratio, good speech intelligibility can be achieved even with a lot of noise attenuation. It can be shown that a traditional pure stats system can attenuate noise better than its modern counterpart because it is older and has a larger sample size. Older classic pure stats systems use less processing power, hence they’re more efficient. We’ve created a new a priori method for measuring SNR for voice-enhancement apps. The cepstral domain spectral envelope is estimated using a DNN. We can improve the quality of low-order noise models while also providing listeners with a natural and pleasurable background noise experience using our CEM approach.